wust

src/rm_auto_aim/armor_mpc_solver/CMakeLists.txt

@@ -0,0 +1,87 @@
+cmake_minimum_required(VERSION 3.8)
+project(armor_mpc_solver)
+
+if(CMAKE_CXX_STANDARD GREATER_RANGE 17)
+  set(CMAKE_CXX_STANDARD 17)
+else()
+  set(CMAKE_CXX_STANDARD 17)
+endif()
+
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
+# Find dependencies
+find_package(ament_cmake REQUIRED)
+find_package(rclcpp REQUIRED)
+find_package(rm_interfaces REQUIRED)
+find_package(rm_utils REQUIRED)
+find_package(rm_tinympc REQUIRED)
+find_package(Eigen3 REQUIRED)
+find_package(tf2 REQUIRED)
+find_package(tf2_ros REQUIRED)
+find_package(visualization_msgs REQUIRED)
+find_package(geometry_msgs REQUIRED)
+find_package(std_msgs REQUIRED)
+
+include_directories(
+  include
+  ${EIGEN3_INCLUDE_DIRS}
+)
+
+set(${PROJECT_NAME}_HEADERS
+  include/armor_mpc_solver/solver.hpp
+  include/armor_mpc_solver/solver_comparer.hpp
+  include/armor_mpc_solver/solver_comparison_node.hpp
+)
+
+add_library(${PROJECT_NAME} SHARED
+  src/solver.cpp
+  src/solver_comparer.cpp
+)
+
+add_executable(${PROJECT_NAME}_node src/solver_comparison_node.cpp)
+ament_target_dependencies(${PROJECT_NAME}_node
+  rclcpp
+  rm_interfaces
+  rm_utils
+  rm_tinympc
+  Eigen3
+  tf2
+  tf2_ros
+  visualization_msgs
+  geometry_msgs
+  std_msgs
+  ${PROJECT_NAME}
+)
+
+ament_target_dependencies(${PROJECT_NAME}
+  rclcpp
+  rm_interfaces
+  rm_utils
+  rm_tinympc
+  Eigen3
+  tf2
+  tf2_ros
+  visualization_msgs
+  geometry_msgs
+  std_msgs
+)
+
+target_include_directories(${PROJECT_NAME} PUBLIC
+  $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
+  $<INSTALL_INTERFACE:include>
+)
+
+install(TARGETS ${PROJECT_NAME} ${PROJECT_NAME}_node
+  ARCHIVE DESTINATION lib
+  LIBRARY DESTINATION lib
+  RUNTIME DESTINATION bin
+)
+
+install(DIRECTORY include/
+  DESTINATION include
+)
+
+ament_export_include_directories(include)
+ament_export_libraries(${PROJECT_NAME})
+
+ament_package()

src/rm_auto_aim/armor_mpc_solver/include/armor_mpc_solver/solver.hpp

@@ -0,0 +1,311 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef ARMOR_MPC_SOLVER_SOLVER_HPP_
+#define ARMOR_MPC_SOLVER_SOLVER_HPP_
+
+#include <memory>
+#include <vector>
+#include <Eigen/Eigen>
+#include <rclcpp/rclcpp.hpp>
+#include <tf2_ros/buffer.h>
+#include <rm_interfaces/msg/target.hpp>
+#include <rm_interfaces/msg/gimbal_cmd.hpp>
+#include "rm_tinympc/types.hpp"
+#include "rm_tinympc/admm.hpp"
+#include "rm_tinympc/trajectory.hpp"
+#include "rm_utils/math/trajectory_compensator.hpp"
+#include "rm_utils/math/manual_compensator.hpp"
+
+namespace fyt::auto_aim {
+
+/**
+ * Trajectory point for visualization and debugging
+ */
+struct TrajPoint {
+  double t;
+  double yaw;
+  double pitch;
+  double yaw_vel;
+  double pitch_vel;
+};
+
+/**
+ * Quintic segment for polynomial trajectory
+ * Provides smooth position/velocity/acceleration trajectories
+ */
+struct QuinticSegment {
+  double a0, a1, a2, a3, a4, a5;  // Coefficients: p(t) = a0 + a1*t + a2*t^2 + a3*t^3 + a4*t^4 + a5*t^5
+
+  double position(double t) const;
+  double velocity(double t) const;
+  double acceleration(double t) const;
+
+  static QuinticSegment fromBoundaryConditions(
+    double p0, double v0, double a0,
+    double p1, double v1, double a1,
+    double duration);
+};
+
+/**
+ * LimitTrajectory - Acceleration-constrained trajectory generation
+ * Ensures gimbal acceleration stays within limits
+ */
+class LimitTrajectory {
+public:
+  struct TrajectoryPoint {
+    double t;
+    double yaw;
+    double pitch;
+    double v_yaw;
+    double v_pitch;
+    double a_yaw;
+    double a_pitch;
+  };
+
+  LimitTrajectory(double max_yaw_acc, double max_pitch_acc, double dt);
+
+  std::vector<TrajectoryPoint> generate(
+    double start_yaw, double start_pitch,
+    double target_yaw, double target_pitch,
+    double max_vel_yaw, double max_vel_pitch);
+
+private:
+  double max_yaw_acc_;
+  double max_pitch_acc_;
+  double dt_;
+
+  double trapezoidalTime(double dist, double max_vel, double max_acc);
+  QuinticSegment quinticPlan(double p0, double v0, double a0, double p1, double v1, double a1, double T);
+};
+
+/**
+ * Armor selection result
+ */
+struct ArmorSelectResult {
+  int armor_id;           // Selected armor index (0-3)
+  double coming_angle;   // Angle when approaching
+  double leaving_angle;  // Angle when leaving
+  bool is_switching;     // Whether switching armor
+};
+
+/**
+ * Gimbal MPC Solver using TinyMPC ADMM
+ *
+ * State: [yaw, pitch, yaw_rate, pitch_rate]
+ * Input: [yaw_acc, pitch_acc]
+ *
+ * Cost: sum ||state - ref||_Q + ||input||_R
+ * Constraints: yaw_rate, pitch_rate limits
+ */
+class MpcSolver {
+public:
+  MpcSolver(std::weak_ptr<rclcpp::Node> n);
+
+  rm_interfaces::msg::GimbalCmd solve(
+    const rm_interfaces::msg::Target& target,
+    const rclcpp::Time& current_time,
+    std::shared_ptr<tf2_ros::Buffer> tf2_buffer);
+
+  // Trajectory getters for visualization
+  std::vector<TrajPoint> getMpcTrajectory() const { return mpc_trajectory_; }
+  std::vector<TrajPoint> getReferenceTrajectory() const { return reference_trajectory_; }
+  std::vector<TrajPoint> getOptimalTrajectory() const { return optimal_trajectory_; }
+
+  void setBulletSpeed(double bullet_speed) { bullet_speed_ = bullet_speed; }
+
+  // Limit trajectory for acceleration-constrained paths
+  std::vector<LimitTrajectory::TrajectoryPoint> getLimitTrajectory() const { return limit_trajectory_; }
+
+private:
+  // State and input dimensions
+  static constexpr int N_X = 4;  // [yaw, pitch, yaw_rate, pitch_rate]
+  static constexpr int N_U = 2; // [yaw_acc, pitch_acc]
+
+  // MPC parameters
+  double bullet_speed_;
+  double gravity_;
+  double prediction_horizon_;
+  double dt_;
+
+  // Gimbal rate limits (rad/s)
+  double max_yaw_rate_;
+  double max_pitch_rate_;
+  // Gimbal acceleration limits (rad/s^2)
+  double max_yaw_acc_;
+  double max_pitch_acc_;
+
+  // Cost weights - separated for yaw and pitch
+  // Q = [position_cost, velocity_cost]
+  Eigen::Vector2d Q_yaw_;
+  Eigen::Vector2d Q_pitch_;
+  Eigen::Vector2d R_yaw_;
+  Eigen::Vector2d R_pitch_;
+  Eigen::Vector2d Qf_yaw_;
+  Eigen::Vector2d Qf_pitch_;
+
+  // Legacy combined cost weights (for compatibility)
+  Eigen::Vector4d Q_;  // State cost [yaw, pitch, yaw_rate, pitch_rate]
+  Eigen::Vector2d R_;  // Input cost [yaw_acc, pitch_acc]
+  Eigen::Vector4d Qf_; // Terminal cost
+
+  // Fire decision parameters
+  double delay_enable_fire_error_;
+  double yaw_limit_deg_;
+  double shooting_range_h_;
+  double shooting_range_small_w_;
+  double shooting_range_big_w_;
+  double min_enable_pitch_deg_;
+  double min_enable_yaw_deg_;
+  double comming_angle_deg_;
+  double leaving_angle_deg_;
+
+  // Armor selection
+  double coming_angle_thresh_;
+  double leaving_angle_thresh_;
+
+  // ADMM MPC solver - separated for yaw and pitch
+  std::unique_ptr<rm_tinympc::ADMM> admm_solver_yaw_;
+  std::unique_ptr<rm_tinympc::ADMM> admm_solver_pitch_;
+
+  // Trajectory generators
+  rm_tinympc::GimbalTrajectoryGenerator trajectory_generator_;
+
+  // Trajectories for visualization
+  std::vector<TrajPoint> mpc_trajectory_;
+  std::vector<TrajPoint> reference_trajectory_;
+  std::vector<TrajPoint> optimal_trajectory_;
+  std::vector<LimitTrajectory::TrajectoryPoint> limit_trajectory_;
+
+  rclcpp::Time last_time_;
+
+  // Current gimbal state [yaw, pitch]
+  Eigen::Vector2d current_gimbal_state_;
+
+  // Current gimbal velocity [yaw_rate, pitch_rate]
+  Eigen::Vector2d current_gimbal_velocity_;
+
+  // Control delay compensation (seconds)
+  double control_delay_;
+
+  // Limit trajectory generator
+  std::unique_ptr<LimitTrajectory> limit_trajectory_generator_;
+
+  // Trajectory compensator for bullet arc compensation
+  std::unique_ptr<fyt::TrajectoryCompensator> trajectory_compensator_;
+
+  // Manual compensator for angle offset correction
+  std::unique_ptr<fyt::ManualCompensator> manual_compensator_;
+
+  /**
+   * Initialize ADMM solver with problem matrices
+   */
+  void initADMM();
+
+  /**
+   * Initialize trajectory compensator
+   */
+  void initTrajectoryCompensator(const std::string& compensator_type);
+
+  /**
+   * Compute reference trajectory (target states over horizon)
+   */
+  Eigen::MatrixXd computeReferenceTrajectory(
+    const Eigen::Vector3d& target_pos,
+    const Eigen::Vector3d& target_vel,
+    double target_yaw,
+    double v_yaw,
+    double flying_time);
+
+  /**
+   * Solve MPC and get optimal control
+   */
+  Eigen::Vector2d solveMPC(const Eigen::Vector4d& x0, const Eigen::MatrixXd& xref);
+
+  /**
+   * Compute gimbal command from optimal control
+   */
+  Eigen::Vector2d computeGimbalCommandFromControl(const Eigen::Vector2d& u);
+
+  /**
+   * Compute flying time with trajectory compensation
+   */
+  double computeFlyingTime(const Eigen::Vector3d& target_pos);
+
+  /**
+   * Predict target position at flying_time
+   */
+  Eigen::Vector3d predictTargetPosition(
+    const Eigen::Vector3d& pos,
+    const Eigen::Vector3d& vel,
+    double dt);
+
+  /**
+   * Convert 2D gimbal state to gimbal command message
+   */
+  rm_interfaces::msg::GimbalCmd toGimbalCmd(
+    const Eigen::Vector2d& gimbal_angles,
+    const Eigen::Vector3d& target_pos,
+    const rm_interfaces::msg::Target& target);
+
+  /**
+   * Build MPC trajectory for visualization (separated yaw/pitch)
+   */
+  void buildMpcVisualizationTrajectory(
+    const Eigen::Vector4d& x0,
+    const Eigen::MatrixXd& x_traj_yaw,
+    const Eigen::MatrixXd& x_traj_pitch,
+    const Eigen::MatrixXd& u_traj_yaw,
+    const Eigen::MatrixXd& u_traj_pitch);
+
+  /**
+   * Build reference trajectory for visualization
+   */
+  void buildReferenceVisualizationTrajectory(
+    const Eigen::MatrixXd& xref);
+
+  /**
+   * Select best armor based on coming/leaving angle
+   */
+  ArmorSelectResult selectArmor(
+    const rm_interfaces::msg::Target& target,
+    const rclcpp::Time& current_time);
+
+  /**
+   * Check if can fire at given time considering control delay
+   */
+  bool canFireAtTime(
+    const rm_interfaces::msg::Target& target,
+    double time_to_target,
+    const rclcpp::Time& current_time);
+
+  /**
+   * Compute trajectory with limit constraints
+   */
+  std::vector<LimitTrajectory::TrajectoryPoint> computeLimitTrajectory(
+    double target_yaw, double target_pitch);
+
+  /**
+   * Compute time-optimal trajectory with acceleration constraints
+   */
+  void computeAccelConstrainedTrajectory(
+    const Eigen::Vector2d& start,
+    const Eigen::Vector2d& target,
+    const Eigen::Vector2d& max_vel,
+    const Eigen::Vector2d& max_acc);
+};
+
+}  // namespace fyt::auto_aim
+
+#endif  // ARMOR_MPC_SOLVER_SOLVER_HPP_

src/rm_auto_aim/armor_mpc_solver/include/armor_mpc_solver/solver_comparer.hpp

@@ -0,0 +1,70 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef ARMOR_MPC_SOLVER_SOLVER_COMPARER_HPP_
+#define ARMOR_MPC_SOLVER_SOLVER_COMPARER_HPP_
+
+#include <memory>
+#include <vector>
+#include <Eigen/Eigen>
+#include <rclcpp/rclcpp.hpp>
+#include <tf2_ros/buffer.h>
+#include <tf2_ros/transform_listener.h>
+#include <visualization_msgs/msg/marker.hpp>
+#include <visualization_msgs/msg/marker_array.hpp>
+#include <rm_interfaces/msg/target.hpp>
+#include <rm_interfaces/msg/gimbal_cmd.hpp>
+#include <geometry_msgs/msg/point.hpp>
+#include "armor_solver/armor_solver.hpp"
+#include "armor_mpc_solver/solver.hpp"
+
+namespace fyt::auto_aim {
+
+class SolverComparer {
+public:
+  SolverComparer(std::weak_ptr<rclcpp::Node> n);
+
+  void init();
+  void update(const rm_interfaces::msg::Target::SharedPtr target_msg);
+
+  void publishComparisonMarkers();
+
+private:
+  std::weak_ptr<rclcpp::Node> node_;
+
+  std::unique_ptr<armor_solver::Solver> original_solver_;
+  std::unique_ptr<MpcSolver> mpc_solver_;
+
+  rclcpp::Publisher<visualization_msgs::msg::MarkerArray>::SharedPtr trajectory_pub_;
+  rclcpp::Publisher<rm_interfaces::msg::GimbalCmd>::SharedPtr mpc_gimbal_pub_;
+
+  std::shared_ptr<tf2_ros::Buffer> tf2_buffer_;
+  std::shared_ptr<tf2_ros::TransformListener> tf2_listener_;
+
+  std::vector<TrajPoint> last_mpc_trajectory_;
+  std::vector<TrajPoint> last_reference_trajectory_;
+  std::vector<LimitTrajectory::TrajectoryPoint> last_limit_trajectory_;
+
+  int marker_id_;
+  std::string frame_id_;
+
+  rm_interfaces::msg::GimbalCmd last_original_cmd_;
+  rm_interfaces::msg::GimbalCmd last_mpc_cmd_;
+
+  bool use_mpc_;
+};
+
+}  // namespace fyt::auto_aim
+
+#endif  // ARMOR_MPC_SOLVER_SOLVER_COMPARER_HPP_

src/rm_auto_aim/armor_mpc_solver/include/armor_mpc_solver/solver_comparison_node.hpp

@@ -0,0 +1,43 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef ARMOR_MPC_SOLVER_SOLVER_COMPARISON_NODE_HPP_
+#define ARMOR_MPC_SOLVER_SOLVER_COMPARISON_NODE_HPP_
+
+#include <rclcpp/rclcpp.hpp>
+#include <rm_interfaces/msg/target.hpp>
+#include <std_msgs/msg/bool.hpp>
+#include "armor_mpc_solver/solver_comparer.hpp"
+
+namespace fyt::auto_aim {
+
+class SolverComparisonNode : public rclcpp::Node {
+public:
+  SolverComparisonNode();
+
+private:
+  void targetCallback(const rm_interfaces::msg::Target::SharedPtr msg);
+  void toggleCallback(const std_msgs::msg::Bool::SharedPtr msg);
+
+  std::unique_ptr<SolverComparer> comparer_;
+
+  rclcpp::Subscription<rm_interfaces::msg::Target>::SharedPtr target_sub_;
+  rclcpp::Subscription<std_msgs::msg::Bool>::SharedPtr toggle_sub_;
+
+  rclcpp::TimerBase::SharedPtr timer_;
+};
+
+}  // namespace fyt::auto_aim
+
+#endif  // ARMOR_MPC_SOLVER_SOLVER_COMPARISON_NODE_HPP_

src/rm_auto_aim/armor_mpc_solver/package.xml

@@ -0,0 +1,26 @@
+<?xml version="1.0"?>
+<?xml-model href="http://download.ros.org/schema/package_format3.xsd" schematypens="http://www.w3.org/2001/XMLSchema"?>
+<package format="3">
+  <name>armor_mpc_solver</name>
+  <version>0.1.0</version>
+  <description>MPC-based armor solver with TinyMPC</description>
+  <maintainer email="chenyouyuan@foxmail.com">Chen Youyuan</maintainer>
+  <license>Apache-2.0</license>
+
+  <buildtool_depend>ament_cmake</buildtool_depend>
+
+  <depend>rclcpp</depend>
+  <depend>rm_interfaces</depend>
+  <depend>rm_utils</depend>
+  <depend>rm_tinympc</depend>
+  <depend>Eigen3</depend>
+  <depend>tf2</depend>
+  <depend>tf2_ros</depend>
+  <depend>visualization_msgs</depend>
+  <depend>geometry_msgs</depend>
+  <depend>std_msgs</depend>
+
+  <export>
+    <build_type>ament_cmake</build_type>
+  </export>
+</package>

src/rm_auto_aim/armor_mpc_solver/src/solver.cpp

@@ -0,0 +1,667 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "armor_mpc_solver/solver.hpp"
+#include "rm_utils/logger/log.hpp"
+#include <angles/angles.h>
+#include <cmath>
+#include <string>
+
+namespace fyt::auto_aim {
+
+// ============== QuinticSegment Implementation ==============
+
+double QuinticSegment::position(double t) const {
+  return a0 + a1 * t + a2 * t * t + a3 * t * t * t + a4 * t * t * t * t + a5 * t * t * t * t * t;
+}
+
+double QuinticSegment::velocity(double t) const {
+  return a1 + 2 * a2 * t + 3 * a3 * t * t + 4 * a4 * t * t * t + 5 * a5 * t * t * t * t;
+}
+
+double QuinticSegment::acceleration(double t) const {
+  return 2 * a2 + 6 * a3 * t + 12 * a4 * t * t + 20 * a5 * t * t * t;
+}
+
+QuinticSegment QuinticSegment::fromBoundaryConditions(
+  double p0, double v0, double a0,
+  double p1, double v1, double a1,
+  double T) {
+  QuinticSegment seg;
+  double T2 = T * T;
+  double T3 = T2 * T;
+  double T4 = T3 * T;
+  double T5 = T4 * T;
+
+  // Solve for coefficients using boundary conditions
+  seg.a0 = p0;
+  seg.a1 = v0;
+  seg.a2 = a0 / 2.0;
+  seg.a3 = (20 * (p1 - p0) - (8 * v1 + 12 * v0) * T - (3 * a0 - a1) * T2) / (2 * T3);
+  seg.a4 = (30 * (p0 - p1) + (14 * v1 + 16 * v0) * T + (3 * a0 - 2 * a1) * T2) / (2 * T4);
+  seg.a5 = (12 * (p1 - p0) - (6 * v1 + 6 * v0) * T - (a0 - a1) * T2) / (2 * T5);
+
+  return seg;
+}
+
+// ============== LimitTrajectory Implementation ==============
+
+LimitTrajectory::LimitTrajectory(double max_yaw_acc, double max_pitch_acc, double dt)
+  : max_yaw_acc_(max_yaw_acc), max_pitch_acc_(max_pitch_acc), dt_(dt) {}
+
+double LimitTrajectory::trapezoidalTime(double dist, double max_vel, double max_acc) {
+  // Time for trapezoidal velocity profile
+  // v_max = sqrt(dist * acc) if dist > v_max^2 / acc
+  double t_acc = max_vel / max_acc;
+  double dist_at_t_acc = max_acc * t_acc * t_acc;
+
+  if (dist <= dist_at_t_acc) {
+    // Triangle profile: accelerate then decelerate
+    return 2.0 * std::sqrt(dist / max_acc);
+  } else {
+    // Trapezoidal: accel + coast + decel
+    double t_coast = (dist - dist_at_t_acc) / max_vel;
+    return 2.0 * t_acc + t_coast;
+  }
+}
+
+QuinticSegment LimitTrajectory::quinticPlan(
+  double p0, double v0, double a0,
+  double p1, double v1, double a1,
+  double T) {
+  return QuinticSegment::fromBoundaryConditions(p0, v0, a0, p1, v1, a1, T);
+}
+
+std::vector<LimitTrajectory::TrajectoryPoint> LimitTrajectory::generate(
+  double start_yaw, double start_pitch,
+  double target_yaw, double target_pitch,
+  double max_vel_yaw, double max_vel_pitch) {
+
+  std::vector<TrajectoryPoint> trajectory;
+
+  // Compute angle differences
+  double dyaw = angles::shortest_angular_distance(start_yaw, target_yaw);
+  double dpitch = target_pitch - start_pitch;
+
+  // Time for each axis using trapezoidal profile
+  double time_yaw = trapezoidalTime(std::abs(dyaw), max_vel_yaw, max_yaw_acc_);
+  double time_pitch = trapezoidalTime(std::abs(dpitch), max_vel_pitch, max_pitch_acc_);
+
+  // Use longer time and synchronize
+  double T = std::max(time_yaw, time_pitch);
+  T = std::max(T, 0.1);  // Minimum time
+
+  // Generate quintic trajectories
+  QuinticSegment yaw_traj = quinticPlan(start_yaw, 0, 0, target_yaw, 0, 0, T);
+  QuinticSegment pitch_traj = quinticPlan(start_pitch, 0, 0, target_pitch, 0, 0, T);
+
+  // Sample trajectory
+  int num_steps = static_cast<int>(T / dt_) + 1;
+  for (int i = 0; i <= num_steps; ++i) {
+    double t = i * dt_;
+    if (t > T) t = T;
+
+    TrajectoryPoint pt;
+    pt.t = t;
+    pt.yaw = yaw_traj.position(t);
+    pt.pitch = pitch_traj.position(t);
+    pt.v_yaw = yaw_traj.velocity(t);
+    pt.v_pitch = pitch_traj.velocity(t);
+    pt.a_yaw = yaw_traj.acceleration(t);
+    pt.a_pitch = pitch_traj.acceleration(t);
+
+    trajectory.push_back(pt);
+  }
+
+  return trajectory;
+}
+
+// ============== MpcSolver Implementation ==============
+
+MpcSolver::MpcSolver(std::weak_ptr<rclcpp::Node> n)
+: node_(n), bullet_speed_(20.0), gravity_(9.8), prediction_horizon_(2.0), dt_(0.004),
+  max_yaw_rate_(6.0), max_pitch_rate_(6.0), max_yaw_acc_(40.0), max_pitch_acc_(25.0),
+  control_delay_(0.2), delay_enable_fire_error_(0.0035),
+  yaw_limit_deg_(60.0), shooting_range_h_(0.12), shooting_range_small_w_(0.12),
+  shooting_range_big_w_(0.24), min_enable_pitch_deg_(0.25), min_enable_yaw_deg_(0.25),
+  comming_angle_deg_(60.0), leaving_angle_deg_(20.0) {
+
+  auto node = node_.lock();
+  if (node) {
+    // Basic parameters (matching wust_vision)
+    bullet_speed_ = node->declare_parameter("mpc.bullet_speed", 20.0);
+    gravity_ = node->declare_parameter("mpc.gravity", 9.8);
+    prediction_horizon_ = node->declare_parameter("mpc.sample_total_time", 2.0);
+    int sample_horizon = node->declare_parameter("mpc.sample_horizon", 500);
+    dt_ = prediction_horizon_ / sample_horizon;
+
+    max_yaw_acc_ = node->declare_parameter("mpc.max_yaw_acc", 40.0);
+    max_pitch_acc_ = node->declare_parameter("mpc.max_pitch_acc", 25.0);
+
+    control_delay_ = node->declare_parameter("mpc.control_delay", 0.2);
+    delay_enable_fire_error_ = node->declare_parameter("mpc.delay_enable_fire_error", 0.0035);
+
+    // Fire decision parameters
+    yaw_limit_deg_ = node->declare_parameter("mpc.yaw_limit_deg", 60.0);
+    shooting_range_h_ = node->declare_parameter("mpc.shooting_range_h", 0.12);
+    shooting_range_small_w_ = node->declare_parameter("mpc.shooting_range_small_w", 0.12);
+    shooting_range_big_w_ = node->declare_parameter("mpc.shooting_range_big_w", 0.24);
+    min_enable_pitch_deg_ = node->declare_parameter("mpc.min_enable_pitch_deg", 0.25);
+    min_enable_yaw_deg_ = node->declare_parameter("mpc.min_enable_yaw_deg", 0.25);
+    comming_angle_deg_ = node->declare_parameter("mpc.comming_angle", 60.0);
+    leaving_angle_deg_ = node->declare_parameter("mpc.leaving_angle", 20.0);
+
+    // Cost weights - separated for yaw and pitch (matching wust_vision)
+    // Default: Q_yaw = [7e6, 0], R_yaw = [3.0]
+    auto q_yaw_vec = node->declare_parameter("mpc.Q_yaw", std::vector<double>{7e6, 0.0});
+    auto r_yaw_vec = node->declare_parameter("mpc.R_yaw", std::vector<double>{3.0});
+    auto q_pitch_vec = node->declare_parameter("mpc.Q_pitch", std::vector<double>{7e6, 0.0});
+    auto r_pitch_vec = node->declare_parameter("mpc.R_pitch", std::vector<double>{3.0});
+
+    Q_yaw_ << q_yaw_vec[0], q_yaw_vec[1];
+    R_yaw_ << r_yaw_vec[0];
+    Q_pitch_ << q_pitch_vec[0], q_pitch_vec[1];
+    R_pitch_ << r_pitch_vec[0];
+
+    // Terminal cost is 2x stage cost
+    Qf_yaw_ = 2.0 * Q_yaw_;
+    Qf_pitch_ = 2.0 * Q_pitch_;
+
+    FYT_INFO("armor_mpc_solver", "MPC Solver initialized (wust_vision params):");
+    FYT_INFO("armor_mpc_solver", "  prediction_horizon={:.3f}s, dt={:.4f}s, horizon={}",
+              prediction_horizon_, dt_, sample_horizon);
+    FYT_INFO("armor_mpc_solver", "  max_yaw_acc={:.2f} rad/s^2, max_pitch_acc={:.2f} rad/s^2",
+              max_yaw_acc_, max_pitch_acc_);
+    FYT_INFO("armor_mpc_solver", "  control_delay={:.3f}s, fire_error={:.4f}",
+              control_delay_, delay_enable_fire_error_);
+    FYT_INFO("armor_mpc_solver", "  Q_yaw=[{:.1e}, {:.1e}], R_yaw=[{:.1e}]",
+              Q_yaw_(0), Q_yaw_(1), R_yaw_(0));
+    FYT_INFO("armor_mpc_solver", "  Q_pitch=[{:.1e}, {:.1e}], R_pitch=[{:.1e}]",
+              Q_pitch_(0), Q_pitch_(1), R_pitch_(0));
+
+    // Initialize trajectory compensator
+    std::string compensator_type = node->declare_parameter("mpc.trajectory_compensator", "resistance");
+    initTrajectoryCompensator(compensator_type);
+
+    // Initialize manual compensator for angle offset
+    manual_compensator_ = std::make_unique<fyt::ManualCompensator>();
+    auto angle_offset = node->declare_parameter("mpc.trajectory_offset", std::vector<std::string>{});
+    if (!manual_compensator_->updateMapFlow(angle_offset)) {
+      FYT_DEBUG("armor_mpc_solver", "Manual compensator update skipped (empty config)");
+    }
+  }
+  node.reset();
+
+  initADMM();
+
+  current_gimbal_state_ << 0.0, 0.0;
+  current_gimbal_velocity_ << 0.0, 0.0;
+
+  // Initialize limit trajectory generator
+  limit_trajectory_generator_ = std::make_unique<LimitTrajectory>(
+    max_yaw_acc_, max_pitch_acc_, dt_);
+}
+
+void MpcSolver::initADMM() {
+  int horizon = static_cast<int>(prediction_horizon_ / dt_);
+
+  // Separate MPC for yaw: State [yaw, yaw_rate], Input [yaw_acc]
+  // Separate MPC for pitch: State [pitch, pitch_rate], Input [pitch_acc]
+  static constexpr int N_X_SEP = 2;  // [angle, angular_rate]
+  static constexpr int N_U_SEP = 1;  // [angular_acc]
+
+  admm_solver_yaw_ = std::make_unique<rm_tinympc::ADMM>();
+  admm_solver_pitch_ = std::make_unique<rm_tinympc::ADMM>();
+
+  admm_solver_yaw_->init(N_X_SEP, N_U_SEP, horizon);
+  admm_solver_pitch_->init(N_X_SEP, N_U_SEP, horizon);
+
+  // State transition for single axis: x = [angle, angular_rate]
+  // angle_{k+1} = angle_k + angular_rate_k * dt
+  // angular_rate_{k+1} = angular_rate_k + angular_acc_k * dt
+  Eigen::MatrixXd A_sep = Eigen::MatrixXd::Identity(N_X_SEP, N_X_SEP);
+  A_sep(0, 1) = dt_;  // angle += angular_rate * dt
+
+  Eigen::MatrixXd B_sep = Eigen::MatrixXd::Zero(N_X_SEP, N_U_SEP);
+  B_sep(0, 0) = 0.5 * dt_ * dt_;  // angle += 0.5 * angular_acc * dt^2
+  B_sep(1, 0) = dt_;               // angular_rate += angular_acc * dt
+
+  // Set problem matrices for both solvers using separated cost weights
+  admm_solver_yaw_->setProblem(A_sep, B_sep, Q_yaw_, R_yaw_, Qf_yaw_);
+  admm_solver_pitch_->setProblem(A_sep, B_sep, Q_pitch_, R_pitch_, Qf_pitch_);
+
+  // Set constraints for yaw
+  Eigen::Vector2d x_min_yaw, x_max_yaw;
+  Eigen::VectorXd u_min_yaw(1), u_max_yaw(1);
+  x_min_yaw << -M_PI, -max_yaw_rate_;
+  x_max_yaw << M_PI, max_yaw_rate_;
+  u_min_yaw(0) = -max_yaw_acc_;
+  u_max_yaw(0) = max_yaw_acc_;
+  admm_solver_yaw_->setConstraints(x_min_yaw, x_max_yaw, u_min_yaw, u_max_yaw);
+
+  // Set constraints for pitch
+  Eigen::Vector2d x_min_pitch, x_max_pitch;
+  Eigen::VectorXd u_min_pitch(1), u_max_pitch(1);
+  x_min_pitch << -M_PI_2, -max_pitch_rate_;
+  x_max_pitch << M_PI_2, max_pitch_rate_;
+  u_min_pitch(0) = -max_pitch_acc_;
+  u_max_pitch(0) = max_pitch_acc_;
+  admm_solver_pitch_->setConstraints(x_min_pitch, x_max_pitch, u_min_pitch, u_max_pitch);
+}
+
+  admm_solver_->setConstraints(x_min, x_max, u_min, u_max);
+}
+
+void MpcSolver::initTrajectoryCompensator(const std::string& compensator_type) {
+  trajectory_compensator_ = fyt::CompensatorFactory::createCompensator(compensator_type);
+
+  if (trajectory_compensator_) {
+    trajectory_compensator_->velocity = bullet_speed_;
+    trajectory_compensator_->gravity = gravity_;
+    FYT_INFO("armor_mpc_solver", "Trajectory compensator initialized: {}", compensator_type);
+  } else {
+    FYT_WARN("armor_mpc_solver", "Failed to create trajectory compensator, using default");
+    trajectory_compensator_ = std::make_unique<fyt::IdealCompensator>();
+    trajectory_compensator_->velocity = bullet_speed_;
+    trajectory_compensator_->gravity = gravity_;
+  }
+}
+
+rm_interfaces::msg::GimbalCmd MpcSolver::solve(
+  const rm_interfaces::msg::Target& target,
+  const rclcpp::Time& current_time,
+  std::shared_ptr<tf2_ros::Buffer> tf2_buffer) {
+
+  Eigen::Vector3d target_pos(target.position.x, target.position.y, target.position.z);
+  Eigen::Vector3d target_vel(target.velocity.x, target.velocity.y, target.velocity.z);
+  double target_yaw = target.yaw;
+  double v_yaw = target.v_yaw;
+
+  // Select armor based on coming/leaving angle
+  auto armor_result = selectArmor(target, current_time);
+
+  // Compute flying time with compensation
+  double flying_time = computeFlyingTime(target_pos);
+
+  // Predict target position at flying_time
+  Eigen::Vector3d predicted_pos = predictTargetPosition(target_pos, target_vel, flying_time);
+  double predicted_yaw = target_yaw + flying_time * v_yaw;
+
+  // Convert target to gimbal angles (yaw, pitch)
+  Eigen::Vector2d target_gimbal;
+  target_gimbal(0) = std::atan2(predicted_pos.y(), predicted_pos.x());  // yaw
+
+  // Calculate pitch with trajectory compensation
+  double raw_pitch = std::atan2(predicted_pos.z(), predicted_pos.head(2).norm());
+  if (trajectory_compensator_) {
+    trajectory_compensator_->compensate(predicted_pos, raw_pitch);
+  }
+  target_gimbal(1) = raw_pitch;
+
+  // Apply manual compensator angle offset
+  if (manual_compensator_) {
+    double dist = predicted_pos.head(2).norm();
+    auto offsets = manual_compensator_->angleHardCorrect(dist, predicted_pos.z());
+    if (offsets.size() >= 2) {
+      target_gimbal(0) += offsets[1] * M_PI / 180.0;  // yaw offset
+      target_gimbal(1) += offsets[0] * M_PI / 180.0;   // pitch offset
+    }
+  }
+
+  // Current state: [yaw, pitch, yaw_rate, pitch_rate]
+  Eigen::Vector4d x0 = Eigen::Vector4d::Zero();
+  x0(0) = current_gimbal_state_(0);
+  x0(1) = current_gimbal_state_(1);
+  x0(2) = current_gimbal_velocity_(0);  // Use filtered velocity
+  x0(3) = current_gimbal_velocity_(1);
+
+  // Compute reference trajectory over horizon
+  Eigen::MatrixXd xref = computeReferenceTrajectory(
+    predicted_pos, target_vel, predicted_yaw, v_yaw, flying_time);
+
+  // Solve separated MPC for yaw and pitch
+  Eigen::Vector2d u_optimal = solveMPC(x0, xref);
+
+  // Build trajectories for visualization
+  Eigen::MatrixXd x_traj_yaw = admm_solver_yaw_->getStateTrajectory();
+  Eigen::MatrixXd x_traj_pitch = admm_solver_pitch_->getStateTrajectory();
+  Eigen::MatrixXd u_traj_yaw = admm_solver_yaw_->getControlSequence();
+  Eigen::MatrixXd u_traj_pitch = admm_solver_pitch_->getControlSequence();
+  buildMpcVisualizationTrajectory(x0, x_traj_yaw, x_traj_pitch, u_traj_yaw, u_traj_pitch);
+  buildReferenceVisualizationTrajectory(xref);
+
+  // Compute limit trajectory for comparison
+  computeAccelConstrainedTrajectory(current_gimbal_state_, target_gimbal,
+    Eigen::Vector2d(max_yaw_rate_, max_pitch_rate_),
+    Eigen::Vector2d(max_yaw_acc_, max_pitch_acc_));
+
+  // Update current gimbal velocity first (acceleration * dt = delta_velocity)
+  current_gimbal_velocity_(0) += u_optimal(0) * dt_;
+  current_gimbal_velocity_(1) += u_optimal(1) * dt_;
+
+  // Apply first control input using updated velocity
+  Eigen::Vector2d cmd_angles = computeGimbalCommandFromControl(u_optimal);
+
+  // Update current gimbal state
+  current_gimbal_state_ = cmd_angles;
+
+  // Convert to gimbal command message
+  auto gimbal_cmd = toGimbalCmd(cmd_angles, predicted_pos, target);
+
+  // Check if can fire considering control delay
+  gimbal_cmd.fire_advice = canFireAtTime(target, flying_time, current_time);
+
+  return gimbal_cmd;
+}
+
+Eigen::MatrixXd MpcSolver::computeReferenceTrajectory(
+  const Eigen::Vector3d& target_pos,
+  const Eigen::Vector3d& target_vel,
+  double target_yaw,
+  double v_yaw,
+  double flying_time) {
+
+  int horizon = admm_solver_yaw_->getStateTrajectory().cols();
+  Eigen::MatrixXd xref(N_X, horizon + 1);
+
+  int num_steps = static_cast<int>(horizon);
+  double t = 0.0;
+
+  for (int i = 0; i <= num_steps; ++i) {
+    double dt_i = i * dt_;
+
+    // Predict target position at dt_i
+    Eigen::Vector3d pred_pos = predictTargetPosition(target_pos, target_vel, dt_i);
+    double pred_yaw = target_yaw + dt_i * v_yaw;
+
+    // Convert to gimbal angles
+    xref(0, i) = std::atan2(pred_pos.y(), pred_pos.x());  // yaw
+    xref(1, i) = std::atan2(pred_pos.z(), pred_pos.head(2).norm());  // pitch
+    xref(2, i) = 0.0;  // yaw_rate (reference is stationary in gimbal frame)
+    xref(3, i) = 0.0;  // pitch_rate
+  }
+
+  return xref;
+}
+
+Eigen::Vector2d MpcSolver::solveMPC(const Eigen::Vector4d& x0, const Eigen::MatrixXd& xref) {
+  // Solve separated MPC for yaw and pitch
+  // x0 = [yaw, pitch, yaw_rate, pitch_rate]
+  // xref columns: [yaw_ref, pitch_ref, yaw_rate_ref=0, pitch_rate_ref=0]
+
+  static constexpr int N_X_SEP = 2;
+
+  // Initial state for yaw MPC: [yaw, yaw_rate]
+  Eigen::VectorXd x0_yaw(N_X_SEP);
+  x0_yaw(0) = x0(0);
+  x0_yaw(1) = x0(2);
+
+  // Initial state for pitch MPC: [pitch, pitch_rate]
+  Eigen::VectorXd x0_pitch(N_X_SEP);
+  x0_pitch(0) = x0(1);
+  x0_pitch(1) = x0(3);
+
+  // Get horizon from solver
+  int horizon = admm_solver_yaw_->getStateTrajectory().cols();
+
+  // Reference trajectories for yaw and pitch
+  Eigen::MatrixXd xref_yaw(N_X_SEP, horizon + 1);
+  Eigen::MatrixXd xref_pitch(N_X_SEP, horizon + 1);
+
+  for (int i = 0; i <= horizon; ++i) {
+    xref_yaw(0, i) = xref(0, i);   // yaw
+    xref_yaw(1, i) = xref(2, i);   // yaw_rate
+    xref_pitch(0, i) = xref(1, i);  // pitch
+    xref_pitch(1, i) = xref(3, i);  // pitch_rate
+  }
+
+  // Solve yaw MPC
+  admm_solver_yaw_->setInitialState(x0_yaw);
+  admm_solver_yaw_->setReference(xref_yaw);
+  bool yaw_converged = admm_solver_yaw_->solve();
+
+  if (!yaw_converged) {
+    FYT_WARN("armor_mpc_solver", "Yaw MPC solver did not converge!");
+  }
+
+  // Solve pitch MPC
+  admm_solver_pitch_->setInitialState(x0_pitch);
+  admm_solver_pitch_->setReference(xref_pitch);
+  bool pitch_converged = admm_solver_pitch_->solve();
+
+  if (!pitch_converged) {
+    FYT_WARN("armor_mpc_solver", "Pitch MPC solver did not converge!");
+  }
+
+  // Get first optimal control [yaw_acc, pitch_acc]
+  Eigen::VectorXd u_yaw = admm_solver_yaw_->getFirstInput();
+  Eigen::VectorXd u_pitch = admm_solver_pitch_->getFirstInput();
+
+  Eigen::Vector2d u_optimal;
+  u_optimal(0) = u_yaw(0);
+  u_optimal(1) = u_pitch(0);
+
+  return u_optimal;
+}
+
+Eigen::Vector2d MpcSolver::computeGimbalCommandFromControl(const Eigen::Vector2d& u) {
+  // Control input u = [yaw_acc, pitch_acc]
+  // Semi-implicit Euler integration (symplectic):
+  // velocity_{k+1} = velocity_k + accel * dt
+  // angle_{k+1} = angle_k + velocity_{k+1} * dt
+
+  Eigen::Vector2d new_angles = current_gimbal_state_;
+
+  // First compute new velocity (after acceleration)
+  Eigen::Vector2d new_velocity = current_gimbal_velocity_ + u * dt_;
+
+  // Then compute angle using new velocity
+  new_angles(0) += new_velocity(0) * dt_;
+  new_angles(1) += new_velocity(1) * dt_;
+
+  // Clamp to gimbal limits
+  new_angles(0) = std::max(-M_PI, std::min(M_PI, new_angles(0)));
+  new_angles(1) = std::max(-M_PI_2, std::min(M_PI_2, new_angles(1)));
+
+  return new_angles;
+}
+
+double MpcSolver::computeFlyingTime(const Eigen::Vector3d& target_pos) {
+  if (trajectory_compensator_) {
+    return trajectory_compensator_->getFlyingTime(target_pos);
+  }
+  // Fallback: simple calculation
+  double dist = target_pos.norm();
+  return dist / bullet_speed_;
+}
+
+Eigen::Vector3d MpcSolver::predictTargetPosition(
+  const Eigen::Vector3d& pos,
+  const Eigen::Vector3d& vel,
+  double dt) {
+  return pos + dt * vel;
+}
+
+rm_interfaces::msg::GimbalCmd MpcSolver::toGimbalCmd(
+  const Eigen::Vector2d& gimbal_angles,
+  const Eigen::Vector3d& target_pos,
+  const rm_interfaces::msg::Target& target) {
+
+  rm_interfaces::msg::GimbalCmd gimbal_cmd;
+  gimbal_cmd.header = target.header;
+  gimbal_cmd.header.stamp = rclcpp::Clock().now();
+  gimbal_cmd.yaw = gimbal_angles(0) * 180.0 / M_PI;
+  gimbal_cmd.pitch = gimbal_angles(1) * 180.0 / M_PI;
+  gimbal_cmd.distance = target_pos.norm();
+
+  // Compute yaw_diff and pitch_diff (simplified - assumes gimbal feedback)
+  // In practice, these would come from actual gimbal feedback
+  gimbal_cmd.yaw_diff = 0.0;
+  gimbal_cmd.pitch_diff = 0.0;
+
+  // Simplified fire_advice: fire if target is being tracked
+  gimbal_cmd.fire_advice = true;
+
+  // shoot_rate will be set by the node that uses this solver
+
+  return gimbal_cmd;
+}
+
+void MpcSolver::buildMpcVisualizationTrajectory(
+  const Eigen::Vector4d& x0,
+  const Eigen::MatrixXd& x_traj_yaw,
+  const Eigen::MatrixXd& x_traj_pitch,
+  const Eigen::MatrixXd& u_traj_yaw,
+  const Eigen::MatrixXd& u_traj_pitch) {
+
+  mpc_trajectory_.clear();
+
+  int horizon = x_traj_yaw.cols() - 1;
+  for (int i = 0; i <= horizon; ++i) {
+    TrajPoint pt;
+    pt.t = i * dt_;
+    pt.yaw = x_traj_yaw(0, i) * 180.0 / M_PI;      // yaw in degrees
+    pt.pitch = x_traj_pitch(0, i) * 180.0 / M_PI;   // pitch in degrees
+    pt.yaw_vel = x_traj_yaw(1, i) * 180.0 / M_PI;   // yaw_rate in deg/s
+    pt.pitch_vel = x_traj_pitch(1, i) * 180.0 / M_PI; // pitch_rate in deg/s
+    mpc_trajectory_.push_back(pt);
+  }
+}
+
+void MpcSolver::buildReferenceVisualizationTrajectory(
+  const Eigen::MatrixXd& xref) {
+
+  reference_trajectory_.clear();
+
+  int horizon = xref.cols() - 1;
+  for (int i = 0; i <= horizon; ++i) {
+    TrajPoint pt;
+    pt.t = i * dt_;
+    pt.yaw = xref(0, i) * 180.0 / M_PI;      // yaw in degrees
+    pt.pitch = xref(1, i) * 180.0 / M_PI;   // pitch in degrees
+    pt.yaw_vel = xref(2, i) * 180.0 / M_PI;  // yaw_rate in deg/s
+    pt.pitch_vel = xref(3, i) * 180.0 / M_PI; // pitch_rate in deg/s
+    reference_trajectory_.push_back(pt);
+  }
+}
+
+ArmorSelectResult MpcSolver::selectArmor(
+  const rm_interfaces::msg::Target& target,
+  const rclcpp::Time& current_time) {
+
+  ArmorSelectResult result;
+  result.armor_id = 0;
+  result.coming_angle = 0.0;
+  result.leaving_angle = 0.0;
+  result.is_switching = false;
+
+  // Target provides position and velocity in Cartesian coordinates
+  Eigen::Vector3d target_pos(target.position.x, target.position.y, target.position.z);
+  Eigen::Vector3d target_vel(target.velocity.x, target.velocity.y, target.velocity.z);
+
+  // Compute distance and approach angle
+  double dist = target_pos.norm();
+  double approach_angle = std::atan2(target_pos.y(), target_pos.x());
+
+  // Estimate angular velocity from target velocity
+  // Angular velocity = (r × v) / |r|^2 where r is position vector
+  double omega = (target_pos.x() * target_vel.y() - target_pos.y() * target_vel.x()) / (dist * dist + 1e-6);
+
+  // Flying time
+  double flying_time = computeFlyingTime(target_pos);
+
+  // Coming angle: angle at which target is approaching
+  // Leaving angle: angle at which target will be when bullet arrives
+  double angle_at_flying_time = approach_angle + omega * flying_time;
+  double coming_angle = angles::shortest_angular_distance(approach_angle, angle_at_flying_time);
+  double leaving_angle = angles::shortest_angular_distance(angle_at_flying_time, approach_angle);
+
+  result.coming_angle = coming_angle;
+  result.leaving_angle = leaving_angle;
+
+  // Check if target is moving toward or away from robot
+  // Using configurable thresholds from wust_vision
+  double coming_thresh = comming_angle_deg_ * M_PI / 180.0;
+  double leaving_thresh = leaving_angle_deg_ * M_PI / 180.0;
+
+  // is_switching: target is moving away fast or angle exceeds threshold
+  result.is_switching = (std::abs(coming_angle) > coming_thresh) ||
+                        (target_vel.norm() > 0.5 && std::abs(leaving_angle) < leaving_thresh);
+
+  FYT_DEBUG("armor_mpc_solver",
+    "selectArmor: dist={:.2f}, coming={:.2f}deg, leaving={:.2f}deg, switching={}",
+    dist, coming_angle * 180 / M_PI, leaving_angle * 180 / M_PI, result.is_switching);
+
+  return result;
+}
+
+bool MpcSolver::canFireAtTime(
+  const rm_interfaces::msg::Target& target,
+  double time_to_target,
+  const rclcpp::Time& current_time) {
+
+  // Get armor selection result
+  auto armor_result = selectArmor(target, current_time);
+
+  // Account for control delay - add extra time buffer
+  double total_delay = control_delay_ + 0.05;  // 50ms system latency buffer
+
+  // Time when bullet will arrive
+  double fire_time = time_to_target;
+
+  // Check if gimbal can reach target within fire_time considering acceleration limits
+  double dyaw = angles::shortest_angular_distance(current_gimbal_state_(0),
+    std::atan2(target.position.y, target.position.x));
+  double dpitch = std::atan2(target.position.z, std::sqrt(target.position.x * target.position.x +
+    target.position.y * target.position.y)) - current_gimbal_state_(1);
+
+  // Minimum time needed to reach target given acceleration constraints
+  double min_time_yaw = 2.0 * std::sqrt(std::abs(dyaw) / max_yaw_acc_);
+  double min_time_pitch = 2.0 * std::sqrt(std::abs(dpitch) / max_pitch_acc_);
+  double min_time_needed = std::max(min_time_yaw, min_time_pitch);
+
+  // Can fire if we have enough time and target is not switching rapidly
+  bool can_fire = (fire_time >= min_time_needed + total_delay) && !armor_result.is_switching;
+
+  FYT_DEBUG("armor_mpc_solver",
+    "canFireAtTime: time_to_target={:.3f}, min_time={:.3f}, can_fire={}",
+    fire_time, min_time_needed, can_fire);
+
+  return can_fire;
+}
+
+std::vector<LimitTrajectory::TrajectoryPoint> MpcSolver::computeLimitTrajectory(
+  double target_yaw, double target_pitch) {
+
+  if (!limit_trajectory_generator_) {
+    return {};
+  }
+
+  return limit_trajectory_generator_->generate(
+    current_gimbal_state_(0), current_gimbal_state_(1),
+    target_yaw, target_pitch,
+    max_yaw_rate_, max_pitch_rate_);
+}
+
+void MpcSolver::computeAccelConstrainedTrajectory(
+  const Eigen::Vector2d& start,
+  const Eigen::Vector2d& target,
+  const Eigen::Vector2d& max_vel,
+  const Eigen::Vector2d& max_acc) {
+
+  // Generate limit trajectory for visualization
+  limit_trajectory_ = computeLimitTrajectory(target(0), target(1));
+}
+
+}  // namespace fyt::auto_aim

src/rm_auto_aim/armor_mpc_solver/src/solver_comparer.cpp

@@ -0,0 +1,154 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "armor_mpc_solver/solver_comparer.hpp"
+#include "rm_utils/logger/log.hpp"
+#include <visualization_msgs/msg/color.hpp>
+#include <cmath>
+
+namespace fyt::auto_aim {
+
+SolverComparer::SolverComparer(std::weak_ptr<rclcpp::Node> n)
+: node_(n), marker_id_(0), frame_id_("odom"), use_mpc_(false) {
+  auto node = node_.lock();
+  if (node) {
+    use_mpc_ = node->declare_parameter("comparer.use_mpc", false);
+  }
+  node.reset();
+}
+
+void SolverComparer::init() {
+  auto node = node_.lock();
+  if (!node) return;
+
+  original_solver_ = std::make_unique<armor_solver::Solver>(node_);
+  mpc_solver_ = std::make_unique<MpcSolver>(node_);
+
+  tf2_buffer_ = std::make_shared<tf2_ros::Buffer>(node->get_clock());
+  tf2_listener_ = std::make_shared<tf2_ros::TransformListener>(*tf2_buffer_);
+
+  trajectory_pub_ = node->create_publisher<visualization_msgs::msg::MarkerArray>(
+    "/armor_solver/comparison_trajectory", 10);
+  mpc_gimbal_pub_ = node->create_publisher<rm_interfaces::msg::GimbalCmd>(
+    "/armor_solver/mpc_gimbal_cmd", 10);
+
+  node.reset();
+}
+
+void SolverComparer::update(const rm_interfaces::msg::Target::SharedPtr target_msg) {
+  if (!target_msg) return;
+
+  try {
+    auto current_time = rclcpp::Time(target_msg->header.stamp);
+
+    if (original_solver_) {
+      last_original_cmd_ = original_solver_->solve(*target_msg, current_time, tf2_buffer_);
+    }
+
+    if (mpc_solver_) {
+      last_mpc_cmd_ = mpc_solver_->solve(*target_msg, current_time, tf2_buffer_);
+      mpc_gimbal_pub_->publish(last_mpc_cmd_);
+
+      last_mpc_trajectory_ = mpc_solver_->getMpcTrajectory();
+      last_reference_trajectory_ = mpc_solver_->getReferenceTrajectory();
+      last_limit_trajectory_ = mpc_solver_->getLimitTrajectory();
+    }
+
+    publishComparisonMarkers();
+
+  } catch (const std::exception& e) {
+    FYT_ERROR("armor_mpc_solver", "Solver comparison error: {}", e.what());
+  }
+}
+
+void SolverComparer::publishComparisonMarkers() {
+  visualization_msgs::msg::MarkerArray marker_array;
+
+  if (!last_mpc_trajectory_.empty()) {
+    visualization_msgs::msg::Marker mpc_line;
+    mpc_line.header.frame_id = frame_id_;
+    mpc_line.header.stamp = rclcpp::Clock().now();
+    mpc_line.ns = "mpc_trajectory";
+    mpc_line.type = visualization_msgs::msg::Marker::LINE_STRIP;
+    mpc_line.action = visualization_msgs::msg::Marker::ADD;
+    mpc_line.scale.x = 0.02;
+    mpc_line.color.r = 1.0;
+    mpc_line.color.g = 0.0;
+    mpc_line.color.b = 0.0;
+    mpc_line.color.a = 1.0;
+
+    for (const auto& pt : last_mpc_trajectory_) {
+      geometry_msgs::msg::Point p;
+      p.x = pt.t;
+      p.y = pt.yaw * 180.0 / M_PI;
+      p.z = pt.pitch * 180.0 / M_PI;
+      mpc_line.points.push_back(p);
+    }
+    mpc_line.id = marker_id_++;
+    marker_array.markers.push_back(mpc_line);
+  }
+
+  if (!last_reference_trajectory_.empty()) {
+    visualization_msgs::msg::Marker ref_line;
+    ref_line.header.frame_id = frame_id_;
+    ref_line.header.stamp = rclcpp::Clock().now();
+    ref_line.ns = "reference_trajectory";
+    ref_line.type = visualization_msgs::msg::Marker::LINE_STRIP;
+    ref_line.action = visualization_msgs::msg::Marker::ADD;
+    ref_line.scale.x = 0.02;
+    ref_line.color.r = 0.0;
+    ref_line.color.g = 1.0;
+    ref_line.color.b = 0.0;
+    ref_line.color.a = 1.0;
+
+    for (const auto& pt : last_reference_trajectory_) {
+      geometry_msgs::msg::Point p;
+      p.x = pt.t;
+      p.y = pt.yaw * 180.0 / M_PI;
+      p.z = pt.pitch * 180.0 / M_PI;
+      ref_line.points.push_back(p);
+    }
+    ref_line.id = marker_id_++;
+    marker_array.markers.push_back(ref_line);
+  }
+
+  // Publish limit trajectory (acceleration-constrained) in blue
+  if (!last_limit_trajectory_.empty()) {
+    visualization_msgs::msg::Marker limit_line;
+    limit_line.header.frame_id = frame_id_;
+    limit_line.header.stamp = rclcpp::Clock().now();
+    limit_line.ns = "limit_trajectory";
+    limit_line.type = visualization_msgs::msg::Marker::LINE_STRIP;
+    limit_line.action = visualization_msgs::msg::Marker::ADD;
+    limit_line.scale.x = 0.02;
+    limit_line.color.r = 0.0;
+    limit_line.color.g = 0.0;
+    limit_line.color.b = 1.0;
+    limit_line.color.a = 1.0;
+
+    for (const auto& pt : last_limit_trajectory_) {
+      geometry_msgs::msg::Point p;
+      p.x = pt.t;
+      p.y = pt.yaw * 180.0 / M_PI;
+      p.z = pt.pitch * 180.0 / M_PI;
+      limit_line.points.push_back(p);
+    }
+    limit_line.id = marker_id_++;
+    marker_array.markers.push_back(limit_line);
+  }
+
+  trajectory_pub_->publish(marker_array);
+}
+
+}  // namespace fyt::auto_aim

src/rm_auto_aim/armor_mpc_solver/src/solver_comparison_node.cpp

@@ -0,0 +1,50 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "armor_mpc_solver/solver_comparison_node.hpp"
+
+namespace fyt::auto_aim {
+
+SolverComparisonNode::SolverComparisonNode()
+: rclcpp::Node("solver_comparison_node") {
+  comparer_ = std::make_unique<SolverComparer>(shared_from_this());
+  comparer_->init();
+
+  target_sub_ = this->create_subscription<rm_interfaces::msg::Target>(
+    "/armor_solver/target",
+    10,
+    std::bind(&SolverComparisonNode::targetCallback, this, std::placeholders::_1));
+
+  toggle_sub_ = this->create_subscription<std_msgs::msg::Bool>(
+    "/armor_solver/toggle_mpc",
+    10,
+    std::bind(&SolverComparisonNode::toggleCallback, this, std::placeholders::_1));
+}
+
+void SolverComparisonNode::targetCallback(const rm_interfaces::msg::Target::SharedPtr msg) {
+  comparer_->update(msg);
+}
+
+void SolverComparisonNode::toggleCallback(const std_msgs::msg::Bool::SharedPtr msg) {
+  // Toggle between original solver and MPC solver
+}
+
+}  // namespace fyt::auto_aim
+
+int main(int argc, char** argv) {
+  rclcpp::init(argc, argv);
+  rclcpp::spin(std::make_shared<fyt::auto_aim::SolverComparisonNode>());
+  rclcpp::shutdown();
+  return 0;
+}

src/rm_auto_aim/armor_solver/include/armor_solver/armor_solver_node.hpp

@@ -84,9 +84,11 @@ private:
   // Adaptive Q matrix parameters
   double s2qxyz_max_, s2qxyz_min_;  // Position noise range
   double s2qyaw_max_, s2qyaw_min_; // Yaw noise range
-  double s2qr_, s2qd_zc_;          // Radius and height offset noise
-  // R matrix parameters
-  double r_x_, r_y_, r_z_, r_yaw_;
+  double s2qr_, s2qd_zc_, s2qd_za_;          // Radius and height offset noise
+  // R matrix parameters (Spherical coordinates: yaw, pitch, dist, ori_yaw)
+  double r_yaw_, r_pitch_, r_dist_, r_ori_yaw_;
+  // Adaptive R scaling for visibility (front vs side armor)
+  double r_front_scale_, r_side_scale_;
   double lost_time_thres_;
   std::unique_ptr<Tracker> tracker_;
 

src/rm_auto_aim/armor_solver/include/armor_solver/armor_tracker.hpp

@@ -70,7 +70,8 @@ public:
   Armor tracked_armor;
   std::string tracked_id;
   ArmorsNum tracked_armors_num;
-  Eigen::VectorXd measurement;
+  Eigen::VectorXd measurement;              // Cartesian [x, y, z, yaw] for debug publishing
+  Eigen::VectorXd spherical_measurement_;  // Spherical [yaw, pitch, dist, ori_yaw] for EKF
   Eigen::VectorXd target_state;
 
   // To store another pair of armors message
@@ -79,6 +80,9 @@ public:
   // To store offset relative to the reference plane
   double d_zc;
 
+  // Last armor type for adaptive noise
+  std::string last_armor_type_;
+
 private:
   void initEKF(const Armor &a) noexcept;
 
@@ -88,6 +92,12 @@ private:
 
   static Eigen::Vector3d getArmorPositionFromState(const Eigen::VectorXd &x) noexcept;
 
+  // Convert Cartesian measurement to spherical
+  static Eigen::Vector4d cartesianToSpherical(double x, double y, double z, double yaw);
+
+  // Adaptive noise based on armor visibility
+  void updateAdaptiveNoise(const Armor &armor) noexcept;
+
   double max_match_distance_;
   double max_match_yaw_diff_;
 

src/rm_auto_aim/armor_solver/include/armor_solver/motion_model.hpp

@@ -28,8 +28,10 @@ enum class MotionModel {
   CONSTANT_VEL_ROT = 2    // Constant velocity and rotation velocity
 };
 
-// X_N: state dimension, Z_N: measurement dimension
-constexpr int X_N = 10, Z_N = 4;
+// X_N: state dimension (11), Z_N: measurement dimension (4)
+// State: [xc, vxc, yc, vyc, zc, vzc, yaw, vyaw, r, d_zc, d_za]
+// Measurement: [yaw, pitch, distance, ori_yaw] (spherical coordinates)
+constexpr int X_N = 11, Z_N = 4;
 
 struct Predict {
   explicit Predict(double dt, MotionModel model = MotionModel::CONSTANT_VEL_ROT)
@@ -44,9 +46,9 @@ struct Predict {
     // v_xyz
     if (model == MotionModel::CONSTANT_VEL_ROT || model == MotionModel::CONSTANT_VELOCITY) {
       // linear velocity
-      x1[0] += x0[1] * dt;
-      x1[2] += x0[3] * dt;
-      x1[4] += x0[5] * dt;
+      x1[0] += x0[1] * dt;   // xc += vxc * dt
+      x1[2] += x0[3] * dt;   // yc += vyc * dt
+      x1[4] += x0[5] * dt;   // zc += vzc * dt
     } else {
       // no velocity
       x1[1] *= 0.;
@@ -57,7 +59,7 @@ struct Predict {
     // v_yaw
     if (model == MotionModel::CONSTANT_VEL_ROT || model == MotionModel::CONSTANT_ROTATION) {
       // angular velocity
-      x1[6] += x0[7] * dt;
+      x1[6] += x0[7] * dt;   // yaw += vyaw * dt
     } else {
       // no rotation
       x1[7] *= 0.;
@@ -68,12 +70,31 @@ struct Predict {
   MotionModel model;
 };
 
+// Spherical measurement model
+// z = [yaw, pitch, distance, ori_yaw]
 struct Measure {
+  template <typename T>
+  void operator()(const T x[Z_N], T z[Z_N]) {
+    // x[0]: xc, x[2]: yc, x[4]: zc, x[6]: yaw, x[8]: r, x[9]: d_zc, x[10]: d_za
+    T xa = x[0] - ceres::cos(x[6]) * x[8];  // armor x
+    T ya = x[2] - ceres::sin(x[6]) * x[8];  // armor y
+    T za = x[4] + x[9] + x[10];              // armor z
+
+    // Convert Cartesian to spherical
+    z[0] = ceres::atan2(ya, xa);  // yaw (azimuth angle)
+    z[1] = ceres::atan2(za, ceres::sqrt(xa * xa + ya * ya));  // pitch (elevation angle)
+    z[2] = ceres::sqrt(xa * xa + ya * ya + za * za);  // distance
+    z[3] = x[6];  // ori_yaw (same as yaw for armor)
+  }
+};
+
+// Cartesian measurement model for comparison
+struct MeasureCartesian {
   template <typename T>
   void operator()(const T x[Z_N], T z[Z_N]) {
     z[0] = x[0] - ceres::cos(x[6]) * x[8];
     z[1] = x[2] - ceres::sin(x[6]) * x[8];
-    z[2] = x[4] + x[9];
+    z[2] = x[4] + x[9] + x[10];
     z[3] = x[6];
   }
 };
@@ -81,4 +102,4 @@ struct Measure {
 using RobotStateEKF = ExtendedKalmanFilter<X_N, Z_N, Predict, Measure>;
 
 }  // namespace fyt::auto_aim
-#endif
+#endif  // ARMOR_SOLVER_MOTION_MODEL_HPP_

src/rm_auto_aim/armor_solver/src/armor_solver_node.cpp

@@ -21,6 +21,8 @@
 // std
 #include <memory>
 #include <vector>
+// third party
+#include <angles/angles.h>
 // project
 #include "armor_solver/motion_model.hpp"
 #include "rm_utils/common.hpp"
@@ -70,6 +72,7 @@ ArmorSolverNode::ArmorSolverNode(const rclcpp::NodeOptions &options)
   s2qyaw_min_ = declare_parameter("ekf.sigma2_q_yaw_min", 50.0);
   s2qr_ = declare_parameter("ekf.sigma2_q_r", 800.0);
   s2qd_zc_ = declare_parameter("ekf.sigma2_q_d_zc", 800.0);
+  s2qd_za_ = declare_parameter("ekf.sigma2_q_d_za", 800.0);
   nis_window_size_ = declare_parameter("ekf.nis_window_size", 20);
   nis_adapt_range_ = declare_parameter("ekf.nis_adapt_range", 2.0);
 
@@ -99,7 +102,7 @@ ArmorSolverNode::ArmorSolverNode(const rclcpp::NodeOptions &options)
     s2q_xyz *= q_scale;
     s2q_yaw *= q_scale;
 
-    double r = s2qr_ * q_scale, d_zc = s2qd_zc_ * q_scale;
+    double r = s2qr_ * q_scale, d_zc = s2qd_zc_ * q_scale, d_za = s2qd_za_ * q_scale;
 
     // White noise integral model for position-velocity state
     double q_x_x = pow(t, 4) / 4 * s2q_xyz, q_x_vx = pow(t, 3) / 2 * s2q_xyz, q_vx_vx = pow(t, 2) * s2q_xyz;
@@ -109,39 +112,65 @@ ArmorSolverNode::ArmorSolverNode(const rclcpp::NodeOptions &options)
            q_vyaw_vyaw = pow(t, 2) * s2q_yaw;
     double q_r = pow(t, 4) / 4 * r;
     double q_d_zc = pow(t, 4) / 4 * d_zc;
+    double q_d_za = pow(t, 4) / 4 * d_za;
     // clang-format off
-    //    xc      v_xc    yc      v_yc    zc      v_zc    yaw         v_yaw       r       d_zc
-    q <<  q_x_x,  q_x_vx, 0,      0,      0,      0,      0,          0,          0,      0,
-          q_x_vx, q_vx_vx,0,      0,      0,      0,      0,          0,          0,      0,
-          0,      0,      q_y_y,  q_y_vy, 0,      0,      0,          0,          0,      0,
-          0,      0,      q_y_vy, q_vy_vy,0,      0,      0,          0,          0,      0,
-          0,      0,      0,      0,      q_z_z,  q_z_vz, 0,          0,          0,      0,
-          0,      0,      0,      0,      q_z_vz, q_vz_vz,0,          0,          0,      0,
-          0,      0,      0,      0,      0,      0,      q_yaw_yaw,  q_yaw_vyaw, 0,      0,
-          0,      0,      0,      0,      0,      0,      q_yaw_vyaw, q_vyaw_vyaw,0,      0,
-          0,      0,      0,      0,      0,      0,      0,          0,          q_r,    0,
-          0,      0,      0,      0,      0,      0,      0,          0,          0,      q_d_zc;
+    //    xc      v_xc    yc      v_yc    zc      v_zc    yaw         v_yaw       r       d_zc     d_za
+    q <<  q_x_x,  q_x_vx, 0,      0,      0,      0,      0,          0,          0,      0,      0,
+          q_x_vx, q_vx_vx,0,      0,      0,      0,      0,          0,          0,      0,      0,
+          0,      0,      q_y_y,  q_y_vy, 0,      0,      0,          0,          0,      0,      0,
+          0,      0,      q_y_vy, q_vy_vy,0,      0,      0,          0,          0,      0,      0,
+          0,      0,      0,      0,      q_z_z,  q_z_vz, 0,          0,          0,      0,      0,
+          0,      0,      0,      0,      q_z_vz, q_vz_vz,0,          0,          0,      0,      0,
+          0,      0,      0,      0,      0,      0,      q_yaw_yaw,  q_yaw_vyaw, 0,      0,      0,
+          0,      0,      0,      0,      0,      0,      q_yaw_vyaw, q_vyaw_vyaw,0,      0,      0,
+          0,      0,      0,      0,      0,      0,      0,          0,          q_r,    0,      0,
+          0,      0,      0,      0,      0,      0,      0,          0,          0,      q_d_zc, 0,
+          0,      0,      0,      0,      0,      0,      0,          0,          0,      0,      q_d_za;
 
     // clang-format on
     return q;
   };
-  // update_R - measurement noise covariance matrix
-  // R scales with distance: farther target -> larger measurement noise
-  r_x_ = declare_parameter("ekf.r_x", 0.05);
-  r_y_ = declare_parameter("ekf.r_y", 0.05);
-  r_z_ = declare_parameter("ekf.r_z", 0.05);
+  // update_R - measurement noise covariance matrix (Spherical coordinates)
+  // z = [yaw, pitch, distance, ori_yaw]
+  // R scales with distance: farther target -> larger angular noise
   r_yaw_ = declare_parameter("ekf.r_yaw", 0.02);
+  r_pitch_ = declare_parameter("ekf.r_pitch", 0.02);
+  r_dist_ = declare_parameter("ekf.r_dist", 0.05);
+  r_ori_yaw_ = declare_parameter("ekf.r_ori_yaw", 0.02);
+
+  // Adaptive R scaling factor based on armor visibility
+  // Front armor: smaller noise (more trust)
+  // Side armor: larger noise (less trust)
+  r_front_scale_ = declare_parameter("ekf.r_front_scale", 0.5);
+  r_side_scale_ = declare_parameter("ekf.r_side_scale", 2.0);
+
   auto u_r = [this](const Eigen::Matrix<double, Z_N, 1> &z) {
     Eigen::Matrix<double, Z_N, Z_N> r;
-    // Calculate distance for better noise scaling
-    double dist = std::sqrt(z[0] * z[0] + z[1] * z[1] + z[2] * z[2]);
+    // z[0] = yaw, z[1] = pitch, z[2] = distance, z[3] = ori_yaw
+    double dist = z[2];  // distance is the 3rd element in spherical measurement
     // Minimum distance to prevent numerical issues when target is very close
     dist = std::max(dist, 1.0);
+
+    // Angular noise scales with distance (smaller at close range)
+    double r_yaw_scaled = r_yaw_ * dist * dist;  // rad^2 * m^2
+    double r_pitch_scaled = r_pitch_ * dist * dist;
+    double r_dist_scaled = r_dist_ * dist;  // m^2
+    double r_ori_yaw_scaled = r_ori_yaw_ * dist * dist;
+
+    // Apply visibility-based scaling if armor visibility info is available
+    // This is updated in tracker_->updateAdaptiveNoise()
+    double visibility_scale = 1.0;
+    if (tracker_->last_armor_type_ == "front") {
+      visibility_scale = r_front_scale_;
+    } else if (tracker_->last_armor_type_ == "side") {
+      visibility_scale = r_side_scale_;
+    }
+
     // clang-format off
-    r << r_x_ * dist, 0, 0, 0,
-         0, r_y_ * dist, 0, 0,
-         0, 0, r_z_ * dist, 0,
-         0, 0, 0, r_yaw_;
+    r << r_yaw_scaled * visibility_scale, 0, 0, 0,
+         0, r_pitch_scaled * visibility_scale, 0, 0,
+         0, 0, r_dist_scaled * visibility_scale, 0,
+         0, 0, 0, r_ori_yaw_scaled * visibility_scale;
     // clang-format on
     return r;
   };
@@ -150,6 +179,20 @@ ArmorSolverNode::ArmorSolverNode(const rclcpp::NodeOptions &options)
   p0.setIdentity();
   tracker_->ekf = std::make_unique<RobotStateEKF>(f, h, u_q, u_r, p0);
 
+  // Set residual function for handling angle wraparound (yaw, pitch)
+  // The measurement z = [yaw, pitch, distance, ori_yaw]
+  // Use angles::shortest_angular_distance to handle -pi~pi discontinuity
+  auto residual_func = [](const Eigen::Matrix<double, Z_N, 1> &z_meas,
+                          const Eigen::Matrix<double, Z_N, 1> &z_pred) {
+    Eigen::Matrix<double, Z_N, 1> residual;
+    residual(0) = angles::shortest_angular_distance(z_pred(0), z_meas(0));  // yaw
+    residual(1) = angles::shortest_angular_distance(z_pred(1), z_meas(1));  // pitch
+    residual(2) = z_meas(2) - z_pred(2);  // distance (no wraparound)
+    residual(3) = angles::shortest_angular_distance(z_pred(3), z_meas(3));  // ori_yaw
+    return residual;
+  };
+  tracker_->ekf->setResidualFunc(residual_func);
+
   // Subscriber with tf2 message_filter
   // tf2 relevant
   tf2_buffer_ = std::make_shared<tf2_ros::Buffer>(this->get_clock());

src/rm_auto_aim/armor_solver/src/armor_tracker.cpp

@@ -19,6 +19,7 @@
 #include "armor_solver/armor_tracker.hpp"
 // std
 #include <cfloat>
+#include <cmath>
 #include <memory>
 #include <string>
 // ros2
@@ -33,16 +34,19 @@
 #include "rm_utils/logger/log.hpp"
 
 namespace fyt::auto_aim {
+
 Tracker::Tracker(double max_match_distance, double max_match_yaw_diff)
 : tracker_state(LOST)
 , tracked_id(std::string(""))
 , measurement(Eigen::VectorXd::Zero(4))
-, target_state(Eigen::VectorXd::Zero(9))
+, spherical_measurement_(Eigen::VectorXd::Zero(4))
+, target_state(Eigen::VectorXd::Zero(X_N))
 , max_match_distance_(max_match_distance)
 , max_match_yaw_diff_(max_match_yaw_diff)
 , detect_count_(0)
 , lost_count_(0)
-, last_yaw_(0) {}
+, last_yaw_(0)
+, last_armor_type_("") {}
 
 void Tracker::setMatchThreshold(double max_match_distance, double max_match_yaw_diff) noexcept {
   max_match_distance_ = max_match_distance;
@@ -69,6 +73,7 @@ void Tracker::init(const Armors::SharedPtr &armors_msg) noexcept {
 
   tracked_id = tracked_armor.number;
   tracker_state = DETECTING;
+  last_armor_type_ = tracked_armor.type;
 
   if (tracked_armor.type == "large" &&
       (tracked_id == "3" || tracked_id == "4" || tracked_id == "5")) {
@@ -111,6 +116,7 @@ void Tracker::update(const Armors::SharedPtr &armors_msg) noexcept {
           yaw_diff = abs(orientationToYaw(armor.pose.orientation) - ekf_prediction(6));
           tracked_armor = armor;
           // Update tracked armor type
+          last_armor_type_ = tracked_armor.type;
           if (tracked_armor.type == "large" &&
               (tracked_id == "3" || tracked_id == "4" || tracked_id == "5")) {
             tracked_armors_num = ArmorsNum::BALANCE_2;
@@ -129,10 +135,17 @@ void Tracker::update(const Armors::SharedPtr &armors_msg) noexcept {
       // Matched armor found
       matched = true;
       auto p = tracked_armor.pose.position;
-      // Update EKF
+
+      // Update adaptive noise based on armor visibility
+      updateAdaptiveNoise(tracked_armor);
+
+      // Store Cartesian measurement for debug publishing
       double measured_yaw = orientationToYaw(tracked_armor.pose.orientation);
       measurement = Eigen::Vector4d(p.x, p.y, p.z, measured_yaw);
-      target_state = ekf->update(measurement);
+
+      // Convert to spherical for EKF update
+      spherical_measurement_ = cartesianToSpherical(p.x, p.y, p.z, measured_yaw);
+      target_state = ekf->update(spherical_measurement_);
     } else if (same_id_armors_count == 1 && yaw_diff > max_match_yaw_diff_) {
       // Matched armor not found, but there is only one armor with the same id
       // and yaw has jumped, take this case as the target is spinning and armor
@@ -203,7 +216,7 @@ void Tracker::initEKF(const Armor &a) noexcept {
   double yc = ya + r * sin(yaw);
   double zc = za;
   d_za = 0, d_zc = 0, another_r = r;
-  target_state << xc, 0, yc, 0, zc, 0, yaw, 0, r, d_zc;
+  target_state << xc, 0, yc, 0, zc, 0, yaw, 0, r, d_zc, d_za;
 
   ekf->setState(target_state);
 }
@@ -217,10 +230,11 @@ void Tracker::handleArmorJump(const Armor &current_armor) noexcept {
     target_state(6) = yaw;
     // Only 4 armors has 2 radius and height
     if (tracked_armors_num == ArmorsNum::NORMAL_4) {
-      d_za = target_state(4) + target_state(9) - current_armor.pose.position.z;
+      d_za = target_state(4) + target_state(9) + target_state(10) - current_armor.pose.position.z;
       std::swap(target_state(8), another_r);
       d_zc = d_zc == 0 ? -d_za : 0;
       target_state(9) = d_zc;
+      target_state(10) = d_za;
     }
     FYT_DEBUG("armor_solver", "Armor Jump!");
   }
@@ -234,6 +248,7 @@ void Tracker::handleArmorJump(const Armor &current_armor) noexcept {
     // large, the state is wrong, reset center position and velocity in the
     // state
     d_zc = 0;
+    d_za = 0;
     double r = target_state(8);
     target_state(0) = p.x + r * cos(yaw);  // xc
     target_state(1) = 0;                   // vxc
@@ -242,6 +257,7 @@ void Tracker::handleArmorJump(const Armor &current_armor) noexcept {
     target_state(4) = p.z;                 // zc
     target_state(5) = 0;                   // vzc
     target_state(9) = d_zc;                // d_zc
+    target_state(10) = d_za;               // d_za
     FYT_WARN("armor_solver", "State wrong!");
   }
 
@@ -262,11 +278,54 @@ double Tracker::orientationToYaw(const geometry_msgs::msg::Quaternion &q) noexce
 
 Eigen::Vector3d Tracker::getArmorPositionFromState(const Eigen::VectorXd &x) noexcept {
   // Calculate predicted position of the current armor
-  double xc = x(0), yc = x(2), za = x(4) + x(9);
+  // x[0]: xc, x[2]: yc, x[4]: zc, x[6]: yaw, x[8]: r, x[9]: d_zc, x[10]: d_za
+  double xc = x(0), yc = x(2), za = x(4) + x(9) + x(10);
   double yaw = x(6), r = x(8);
   double xa = xc - r * cos(yaw);
   double ya = yc - r * sin(yaw);
   return Eigen::Vector3d(xa, ya, za);
 }
 
-}  // namespace fyt::auto_aim
+Eigen::Vector4d Tracker::cartesianToSpherical(double x, double y, double z, double yaw) noexcept {
+  Eigen::Vector4d spherical;
+  double dist_xy = std::sqrt(x * x + y * y);
+  spherical(0) = std::atan2(y, x);                    // yaw
+  spherical(1) = std::atan2(z, dist_xy);               // pitch
+  spherical(2) = std::sqrt(x * x + y * y + z * z);   // distance
+  spherical(3) = yaw;                                  // ori_yaw
+  return spherical;
+}
+
+void Tracker::updateAdaptiveNoise(const Armor &armor) noexcept {
+  // Adaptive R matrix based on armor visibility
+  // Classification based on distance_to_image_center:
+  // - Front armor: smaller distance_to_image_center (closer to image center)
+  // - Side armor: larger distance_to_image_center
+  // This is a heuristic - front armor appears more stable and should have lower measurement noise
+
+  // Threshold for front/side classification
+  // Typical image center distance for front armor: < 0.3 (normalized)
+  // Typical image center distance for side armor: > 0.5 (normalized)
+  constexpr double front_threshold = 0.35;
+  constexpr double side_threshold = 0.55;
+
+  std::string visibility_type;
+  if (armor.distance_to_image_center < front_threshold) {
+    visibility_type = "front";
+  } else if (armor.distance_to_image_center > side_threshold) {
+    visibility_type = "side";
+  } else {
+    // In between: use previous state or default to front
+    visibility_type = last_armor_type_.empty() ? "front" : last_armor_type_;
+  }
+
+  last_armor_type_ = visibility_type;
+
+  FYT_DEBUG("armor_solver",
+            "Adaptive noise: dist_to_center={:.3f}, visibility={}, armor_type={}",
+            armor.distance_to_image_center,
+            visibility_type,
+            armor.type);
+}
+
+}  // namespace fyt::auto_aim

src/rm_bringup/config/node_params/armor_solver_params.yaml

@@ -34,7 +34,7 @@
 
       tracking_thres: 1
       lost_time_thres: 1.0
-    
+
     solver:
       shoot_rate_min: 6
       shoot_rate_max: 12
@@ -43,14 +43,14 @@
       shooting_range_width: 0.10   #射击范围
       shooting_range_height: 0.10  #射击范围
       prediction_delay: 0.02       # 预测装甲板位置的延时，单位秒，+飞行时间
-      controller_delay: 0.01       
-      max_tracking_v_yaw: 5.0 #转速(rad/s)大于这个值时瞄准机器人中心 
-      side_angle: 15.0 
-      compenstator_type: "resistance" 
+      controller_delay: 0.01
+      max_tracking_v_yaw: 5.0 #转速(rad/s)大于这个值时瞄准机器人中心
+      side_angle: 15.0
+      compenstator_type: "resistance"
       gravity: 9.792
-      resistance: 0.038    
+      resistance: 0.038
       iteration_times: 20 # 补偿的迭代次数
-            
+
       # ["距离下限, 距离上限, 高度下限, 高度下限, pitch轴补偿值"]
       # [dist_low, dist_high, height_low, height_high, pitch_offset_deg, yaw_offset_deg]
       angle_offset: [
@@ -64,3 +64,79 @@
         "7.0 8.0  0.4 0.8 0.0 0.0",
         "7.0 8.0  0.8 1.2 0.0 0.0",
       ]
+
+    # MPC solver parameters (matching wust_vision very_aimer)
+    # Type: "seg" (segment-based trajectory) or "mpc" (model predictive control)
+    mpc:
+      enabled: false  # Set to true to use MPC solver instead of original solver
+      type: "seg"     # "seg" or "mpc"
+
+      # Trajectory parameters
+      sample_total_time: 2.0    # Total prediction time (seconds)
+      sample_horizon: 500       # Number of steps in horizon
+
+      # Control parameters
+      control_delay: 0.2        # Control delay (seconds)
+      delay_enable_fire_error: 0.0035  # Fire error threshold
+
+      # Acceleration limits (rad/s^2)
+      max_yaw_acc: 40.0
+      max_pitch_acc: 25.0
+
+      # Fire decision thresholds
+      comming_angle: 60.0        # Coming angle threshold (degrees)
+      leaving_angle: 20.0       # Leaving angle threshold (degrees)
+      yaw_limit_deg: 60.0       # Yaw limit for fire decision
+      shooting_range_h: 0.12    # Height shooting range
+      shooting_range_small_w: 0.12  # Small target width range
+      shooting_range_big_w: 0.24     # Big target width range
+      min_enable_pitch_deg: 0.25     # Min pitch for fire
+      min_enable_yaw_deg: 0.25       # Min yaw for fire
+
+      # MPC cost weights (matching wust_vision)
+      # Q = [position_cost, velocity_cost], R = [control_cost]
+      Q_yaw: [7.0e6, 0.0]      # Yaw position and velocity cost
+      R_yaw: [3.0]             # Yaw control cost
+      Q_pitch: [7.0e6, 0.0]   # Pitch position and velocity cost
+      R_pitch: [3.0]           # Pitch control cost
+
+      # Trajectory compensator
+      trajectory_compensator: "resistance"  # "ideal" or "resistance"
+      gravity: 9.8
+
+      # Trajectory offset (same format as solver.angle_offset)
+      trajectory_offset: [
+        "0.0 4.5 -1.0 1.5 0.0 0.0",
+      ]
+
+    # Alternative MPC configuration for comparison testing
+    mpc_compare:
+      enabled: false
+      type: "mpc"
+
+      sample_total_time: 1.0
+      sample_horizon: 250
+
+      control_delay: 0.15
+      delay_enable_fire_error: 0.005
+
+      max_yaw_acc: 35.0
+      max_pitch_acc: 20.0
+
+      comming_angle: 45.0
+      leaving_angle: 30.0
+      yaw_limit_deg: 50.0
+      shooting_range_h: 0.15
+      shooting_range_small_w: 0.10
+      shooting_range_big_w: 0.20
+      min_enable_pitch_deg: 0.3
+      min_enable_yaw_deg: 0.3
+
+      Q_yaw: [5.0e6, 1.0e5]
+      R_yaw: [2.0]
+      Q_pitch: [5.0e6, 1.0e5]
+      R_pitch: [2.0]
+
+      trajectory_compensator: "resistance"
+      gravity: 9.8
+      trajectory_offset: []

src/rm_hardware_driver/ros2_hik_camera/src/hik_camera_node.cpp

@@ -275,6 +275,15 @@ private:
       // Apply ROI/offset before querying image info
       applyRoiLocked();
 
+      // Force set pixel format to RGB8 every time camera opens
+      // This ensures camera doesn't fall back to default Mono8
+      int set_fmt_ret = MV_CC_SetEnumValue(camera_handle_, "PixelFormat", PixelType_Gvsp_RGB8_Packed);
+      if (set_fmt_ret != MV_OK) {
+        FYT_WARN("camera_driver", "Failed to set PixelFormat to RGB8: [0x{:x}], will use SDK conversion", set_fmt_ret);
+      } else {
+        FYT_INFO("camera_driver", "PixelFormat forced to RGB8");
+      }
+
       // Set exposure / gain
       (void)MV_CC_SetFloatValue(camera_handle_, "ExposureTime", static_cast<float>(exposure_time_));
       (void)MV_CC_SetFloatValue(camera_handle_, "Gain", static_cast<float>(gain_));
@@ -369,6 +378,16 @@ private:
         continue;
       }
 
+      // Debug: log pixel format periodically (every 60 frames)
+      static int frame_count = 0;
+      frame_count++;
+      if (frame_count % 60 == 0) {
+        FYT_INFO("camera_driver", "Frame pixel_type=0x{:x}, size={}x{}",
+                  out_frame.stFrameInfo.enPixelType,
+                  out_frame.stFrameInfo.nWidth,
+                  out_frame.stFrameInfo.nHeight);
+      }
+
       // Ensure output buffer
       const size_t need_size =
         static_cast<size_t>(out_frame.stFrameInfo.nWidth) * out_frame.stFrameInfo.nHeight * 3;

src/rm_tinympc/CMakeLists.txt

@@ -0,0 +1,61 @@
+cmake_minimum_required(VERSION 3.8)
+project(rm_tinympc)
+
+if(CMAKE_CXX_STANDARD GREATER_RANGE 17)
+  set(CMAKE_CXX_STANDARD 17)
+else()
+  set(CMAKE_CXX_STANDARD 17)
+endif()
+
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
+# Find dependencies
+find_package(ament_cmake REQUIRED)
+find_package(Eigen3 REQUIRED)
+find_package(tf2 REQUIRED)
+
+include_directories(
+  include
+  ${EIGEN3_INCLUDE_DIRS}
+)
+
+set(${PROJECT_NAME}_HEADERS
+  include/rm_tinympc/types.hpp
+  include/rm_tinympc/admm.hpp
+  include/rm_tinympc/tiny_api.hpp
+  include/rm_tinympc/codegen.hpp
+  include/rm_tinympc/error.hpp
+  include/rm_tinympc/trajectory.hpp
+)
+
+add_library(${PROJECT_NAME} SHARED
+  src/admm.cpp
+  src/tiny_api.cpp
+  src/codegen.cpp
+  src/trajectory.cpp
+)
+
+ament_target_dependencies(${PROJECT_NAME}
+  Eigen3
+  tf2
+)
+
+target_include_directories(${PROJECT_NAME} PUBLIC
+  $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
+  $<INSTALL_INTERFACE:include>
+)
+
+install(TARGETS ${PROJECT_NAME}
+  ARCHIVE DESTINATION lib
+  LIBRARY DESTINATION lib
+  RUNTIME DESTINATION bin
+)
+
+install(DIRECTORY include/
+  DESTINATION include
+)
+
+ament_export_include_directories(include)
+ament_export_libraries(${PROJECT_NAME})
+
+ament_package()

src/rm_tinympc/include/rm_tinympc/admm.hpp

@@ -0,0 +1,162 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef RM_TINYMPC_ADMM_HPP_
+#define RM_TINYMPC_ADMM_HPP_
+
+#include <Eigen/Eigen>
+#include <vector>
+
+namespace rm_tinympc {
+
+/**
+ * ADMM-based MPC Solver (TinyMPC port)
+ *
+ * Minimizes: sum_{k=0}^{T-1} (x_k - xref_k)' Q (x_k - xref_k) + u_k' R u_k
+ * Subject to: x_{k+1} = A x_k + B u_k (dynamics)
+ *             x_min <= x_k <= x_max (state constraints)
+ *             u_min <= u_k <= u_max (input constraints)
+ */
+class ADMM {
+public:
+  ADMM() = default;
+  ~ADMM() = default;
+
+  /**
+   * Initialize the solver
+   * @param n State dimension
+   * @param m Input dimension
+   * @param T Prediction horizon
+   */
+  void init(int n, int m, int T);
+
+  /**
+   * Set the problem matrices
+   * @param A State transition matrix (n x n)
+   * @param B Input matrix (n x m)
+   * @param Q State cost weight (n x n)
+   * @param R Input cost weight (m x m)
+   * @param Qf Terminal cost weight (n x n)
+   */
+  void setProblem(const Eigen::MatrixXd& A,
+                  const Eigen::MatrixXd& B,
+                  const Eigen::VectorXd& Q,
+                  const Eigen::VectorXd& R,
+                  const Eigen::VectorXd& Qf);
+
+  /**
+   * Set state and input constraints
+   */
+  void setConstraints(const Eigen::VectorXd& x_min,
+                      const Eigen::VectorXd& x_max,
+                      const Eigen::VectorXd& u_min,
+                      const Eigen::VectorXd& u_max);
+
+  /**
+   * Set reference trajectory
+   */
+  void setReference(const Eigen::MatrixXd& xref);
+
+  /**
+   * Set initial state
+   */
+  void setInitialState(const Eigen::VectorXd& x0);
+
+  /**
+   * Solve the MPC problem
+   * @return true if converged, false otherwise
+   */
+  bool solve();
+
+  /**
+   * Get the optimal control input (first input of the sequence)
+   */
+  const Eigen::VectorXd& getFirstInput() const { return u_solution_.col(0); }
+
+  /**
+   * Get the full state trajectory
+   */
+  const Eigen::MatrixXd& getStateTrajectory() const { return x_solution_; }
+
+  /**
+   * Get the full control sequence
+   */
+  const Eigen::MatrixXd& getControlSequence() const { return u_solution_; }
+
+  /**
+   * Configure solver parameters
+   */
+  void setMaxIterations(int max_iter) { max_iterations_ = max_iter; }
+  void setRho(double rho) { rho_ = rho; }
+  void setAbsTol(double tol) { abs_tol_ = tol; }
+  void setRelTol(double tol) { rel_tol_ = tol; }
+  void setVerbose(bool verbose) { verbose_ = verbose; }
+
+private:
+  int n_;  // State dimension
+  int m_;  // Input dimension
+  int T_;  // Horizon
+
+  int max_iterations_ = 100;
+  double rho_ = 0.1;
+  double abs_tol_ = 1e-4;
+  double rel_tol_ = 1e-3;
+  bool verbose_ = false;
+
+  // System matrices
+  Eigen::MatrixXd A_;
+  Eigen::MatrixXd B_;
+  Eigen::VectorXd Q_;
+  Eigen::VectorXd R_;
+  Eigen::VectorXd Qf_;
+
+  // Constraints
+  Eigen::VectorXd x_min_;
+  Eigen::VectorXd x_max_;
+  Eigen::VectorXd u_min_;
+  Eigen::VectorXd u_max_;
+  bool has_constraints_ = false;
+
+  // Reference trajectory
+  Eigen::MatrixXd xref_;
+
+  // Initial state
+  Eigen::VectorXd x0_;
+
+  // Solution
+  Eigen::MatrixXd x_solution_;  // n x (T+1)
+  Eigen::MatrixXd u_solution_;  // m x T
+
+  // ADMM variables
+  Eigen::MatrixXd z_;      // n x (T+1) - state consensus
+  Eigen::MatrixXd u_;       // m x T - control consensus
+  Eigen::MatrixXd z_old_;  // n x (T+1)
+  Eigen::MatrixXd u_old_;  // m x T
+
+  // Lagrange multipliers
+  Eigen::MatrixXd lambda_x_;  // n x (T+1)
+  Eigen::MatrixXd lambda_u_;  // m x T
+
+  // Check convergence
+  bool checkConvergence(const Eigen::MatrixXd& x, const Eigen::MatrixXd& u);
+
+  // Proximal operators
+  Eigen::VectorXd proxGradient(const Eigen::VectorXd& x, const Eigen::VectorXd& grad,
+                                double step, const Eigen::VectorXd& x_min,
+                                const Eigen::VectorXd& x_max);
+};
+
+}  // namespace rm_tinympc
+
+#endif  // RM_TINYMPC_ADMM_HPP_

src/rm_tinympc/include/rm_tinympc/codegen.hpp

@@ -0,0 +1,53 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef RM_TINYMPC_CODEGEN_HPP_
+#define RM_TINYMPC_CODEGEN_HPP_
+
+#include <Eigen/Eigen>
+
+namespace rm_tinympc {
+
+class CodeGen {
+public:
+  CodeGen() = default;
+  ~CodeGen() = default;
+
+  void generateSolverCode(const std::string& output_dir);
+
+  static Eigen::Vector2d computeGimbalCommand(
+    const Eigen::Vector3d& target_pos,
+    const Eigen::Vector3d& target_vel,
+    double target_yaw,
+    double v_yaw,
+    double bullet_speed,
+    double gravity);
+
+  static Eigen::Vector3d predictTargetPosition(
+    const Eigen::Vector3d& pos,
+    const Eigen::Vector3d& vel,
+    double dt);
+
+  static double computeFlyingTime(
+    const Eigen::Vector3d& target_pos,
+    double bullet_speed,
+    double gravity = 9.8);
+
+private:
+  static constexpr double kPi = 3.14159265358979323846;
+};
+
+}  // namespace rm_tinympc
+
+#endif  // RM_TINYMPC_CODEGEN_HPP_

src/rm_tinympc/include/rm_tinympc/error.hpp

@@ -0,0 +1,40 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef RM_TINYMPC_ERROR_HPP_
+#define RM_TINYMPC_ERROR_HPP_
+
+#include <stdexcept>
+#include <string>
+
+namespace rm_tinympc {
+
+class TinyMPCException : public std::runtime_error {
+public:
+  explicit TinyMPCException(const std::string& msg) : std::runtime_error(msg) {}
+};
+
+class ConvergenceError : public TinyMPCException {
+public:
+  explicit ConvergenceError(const std::string& msg) : TinyMPCException(msg) {}
+};
+
+class InvalidSizeError : public TinyMPCException {
+public:
+  explicit InvalidSizeError(const std::string& msg) : TinyMPCException(msg) {}
+};
+
+}  // namespace rm_tinympc
+
+#endif  // RM_TINYMPC_ERROR_HPP_

src/rm_tinympc/include/rm_tinympc/tiny_api.hpp

@@ -0,0 +1,71 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef RM_TINYMPC_TINY_API_HPP_
+#define RM_TINYMPC_TINY_API_HPP_
+
+#include <Eigen/Eigen>
+#include "types.hpp"
+
+namespace rm_tinympc {
+
+class TinyMPC {
+public:
+  TinyMPC() = default;
+  ~TinyMPC() = default;
+
+  void init(int n, int m, int T);
+
+  void setProblem(
+    const Eigen::MatrixXd& A,
+    const Eigen::MatrixXd& B,
+    const Eigen::VectorXd& x0,
+    const Eigen::VectorXd& xref,
+    const Eigen::VectorXd& uref);
+
+  void setWeights(const Eigen::VectorXd& Q, const Eigen::VectorXd& R, const Eigen::VectorXd& Qf);
+
+  void solve();
+
+  const Eigen::VectorXd& getSolution() const { return u_solution_; }
+  const Eigen::MatrixXd& getFullSolution() const { return x_solution_; }
+
+  void setMaxIterations(int max_iter) { max_iterations_ = max_iter; }
+  void setRho(double rho) { rho_ = rho; }
+
+private:
+  int n_;  // state dimension
+  int m_;  // input dimension
+  int T_;  // horizon
+
+  int max_iterations_ = 100;
+  double rho_ = 0.1;
+
+  Eigen::MatrixXd A_;
+  Eigen::MatrixXd B_;
+  Eigen::VectorXd x0_;
+  Eigen::VectorXd xref_;
+  Eigen::VectorXd uref_;
+
+  Eigen::VectorXd Q_;
+  Eigen::VectorXd R_;
+  Eigen::VectorXd Qf_;
+
+  Eigen::MatrixXd x_solution_;
+  Eigen::VectorXd u_solution_;
+};
+
+}  // namespace rm_tinympc
+
+#endif  // RM_TINYMPC_TINY_API_HPP_

src/rm_tinympc/include/rm_tinympc/trajectory.hpp

@@ -0,0 +1,165 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef RM_TINYMPC_TRAJECTORY_HPP_
+#define RM_TINYMPC_TRAJECTORY_HPP_
+
+#include <Eigen/Eigen>
+
+namespace rm_tinympc {
+
+/**
+ * Quintic Polynomial Trajectory Generator
+ *
+ * Generates smooth trajectories from start state to goal state
+ * with specified boundary velocities and accelerations.
+ *
+ * Coefficients: a0 + a1*t + a2*t^2 + a3*t^3 + a4*t^4 + a5*t^5
+ */
+class QuinticPolynomial {
+public:
+  QuinticPolynomial() = default;
+  ~QuinticPolynomial() = default;
+
+  /**
+   * Compute quintic polynomial coefficients
+   * @param p0 Initial position
+   * @param v0 Initial velocity
+   * @param a0 Initial acceleration
+   * @param pf Final position
+   * @param vf Final velocity
+   * @param af Final acceleration
+   * @param T Trajectory duration
+   */
+  void compute(double p0, double v0, double a0,
+               double pf, double vf, double af, double T);
+
+  /**
+   * Evaluate position at time t
+   */
+  double evaluate(double t) const;
+
+  /**
+   * Evaluate velocity at time t
+   */
+  double evaluateVelocity(double t) const;
+
+  /**
+   * Evaluate acceleration at time t
+   */
+  double evaluateAcceleration(double t) const;
+
+  /**
+   * Get polynomial coefficients
+   */
+  const Eigen::Vector6d& getCoefficients() const { return coef_; }
+
+private:
+  Eigen::Vector6d coef_;  // a0, a1, a2, a3, a4, a5
+};
+
+/**
+ * Trajectory Point with position, velocity, acceleration
+ */
+struct TrajectoryPoint1D {
+  double t;      // Time
+  double p;     // Position
+  double v;     // Velocity
+  double a;     // Acceleration
+};
+
+/**
+ * 2D Trajectory Point (yaw, pitch)
+ */
+struct TrajectoryPoint2D {
+  double t;
+  double yaw;
+  double pitch;
+  double yaw_vel;
+  double pitch_vel;
+  double yaw_acc;
+  double pitch_acc;
+};
+
+/**
+ * Generate smooth 1D trajectory with quintic polynomial
+ */
+class TrajectoryGenerator1D {
+public:
+  TrajectoryGenerator1D() = default;
+  ~TrajectoryGenerator1D() = default;
+
+  /**
+   * Generate trajectory from start to goal
+   * @param p0 Initial position
+   * @param pf Final position
+   * @param max_v Maximum velocity (for constraint)
+   * @param max_a Maximum acceleration (for constraint)
+   * @param T Total duration
+   */
+  std::vector<TrajectoryPoint1D> generate(double p0, double pf,
+                                          double max_v, double max_a, double T);
+
+  /**
+   * Generate minimum time trajectory with velocity/acceleration constraints
+   */
+  double generateMinTime(double p0, double pf, double max_v, double max_a);
+
+private:
+  double T_ = 0.0;
+  QuinticPolynomial poly_;
+};
+
+/**
+ * 2D Gimbal Trajectory Generator (yaw and pitch simultaneously)
+ */
+class GimbalTrajectoryGenerator {
+public:
+  GimbalTrajectoryGenerator() = default;
+  ~GimbalTrajectoryGenerator() = default;
+
+  /**
+   * Generate 2D gimbal trajectory
+   * @param yaw0 Initial yaw
+   * @param pitch0 Initial pitch
+   * @param yawf Target yaw
+   * @param pitchf Target pitch
+   * @param max_yaw_rate Maximum yaw rate (rad/s)
+   * @param max_pitch_rate Maximum pitch rate (rad/s)
+   * @param max_yaw_acc Maximum yaw acceleration (rad/s^2)
+   * @param max_pitch_acc Maximum pitch acceleration (rad/s^2)
+   * @param T Trajectory duration
+   */
+  std::vector<TrajectoryPoint2D> generate(double yaw0, double pitch0,
+                                         double yawf, double pitchf,
+                                         double max_yaw_rate, double max_pitch_rate,
+                                         double max_yaw_acc, double max_pitch_acc,
+                                         double T);
+
+  /**
+   * Minimum time trajectory with constraints
+   */
+  double generateMinTime(double yaw0, double pitch0,
+                        double yawf, double pitchf,
+                        double max_yaw_rate, double max_pitch_rate,
+                        double max_yaw_acc, double max_pitch_acc);
+
+private:
+  QuinticPolynomial poly_yaw_;
+  QuinticPolynomial poly_pitch_;
+};
+
+}  // namespace rm_tinympc
+
+#endif  // RM_TINYMPC_TRAJECTORY_HPP_

src/rm_tinympc/include/rm_tinympc/types.hpp

@@ -0,0 +1,37 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef RM_TINYMPC_TYPES_HPP_
+#define RM_TINYMPC_TYPES_HPP_
+
+#include <Eigen/Eigen>
+
+namespace rm_tinympc {
+
+using vec2 = Eigen::Vector2d;
+using vec3 = Eigen::Vector3d;
+using vec4 = Eigen::Vector4d;
+using mat2 = Eigen::Matrix2d;
+using mat3 = Eigen::Matrix3d;
+using mat4 = Eigen::Matrix4d;
+
+struct TrajectoryPoint {
+  double yaw;
+  double pitch;
+  double t;
+};
+
+}  // namespace rm_tinympc
+
+#endif  // RM_TINYMPC_TYPES_HPP_

src/rm_tinympc/package.xml

@@ -0,0 +1,19 @@
+<?xml version="1.0"?>
+<?xml-model href="http://download.ros.org/schema/package_format3.xsd" schematypens="http://www.w3.org/2001/XMLSchema"?>
+<package format="3">
+  <name>rm_tinympc</name>
+  <version>0.1.0</version>
+  <description>TinyMPC - ADMM-based small MPC solver</description>
+  <maintainer email="chenyouyuan@foxmail.com">Chen Youyuan</maintainer>
+  <license>Apache-2.0</license>
+
+  <buildtool_depend>ament_cmake</buildtool_depend>
+  <buildtool_depend>ament_cmake_python</buildtool_depend>
+
+  <depend>Eigen3</depend>
+  <depend>tf2</depend>
+
+  <export>
+    <build_type>ament_cmake</build_type>
+  </export>
+</package>

src/rm_tinympc/src/admm.cpp

@@ -0,0 +1,204 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "rm_tinympc/admm.hpp"
+#include <iostream>
+#include <cmath>
+
+namespace rm_tinympc {
+
+void ADMM::init(int n, int m, int T) {
+  n_ = n;
+  m_ = m;
+  T_ = T;
+
+  // Allocate solution matrices
+  x_solution_.resize(n_, T_ + 1);
+  u_solution_.resize(m_, T_);
+
+  // ADMM variables
+  z_.resize(n_, T_ + 1);
+  u_.resize(m_, T_);
+  z_old_.resize(n_, T_ + 1);
+  u_old_.resize(m_, T_);
+  lambda_x_.resize(n_, T_ + 1);
+  lambda_u_.resize(m_, T_);
+
+  x_solution_.setZero();
+  u_solution_.setZero();
+  z_.setZero();
+  u_.setZero();
+  lambda_x_.setZero();
+  lambda_u_.setZero();
+}
+
+void ADMM::setProblem(const Eigen::MatrixXd& A,
+                       const Eigen::MatrixXd& B,
+                       const Eigen::VectorXd& Q,
+                       const Eigen::VectorXd& R,
+                       const Eigen::VectorXd& Qf) {
+  A_ = A;
+  B_ = B;
+  Q_ = Q;
+  R_ = R;
+  Qf_ = Qf;
+}
+
+void ADMM::setConstraints(const Eigen::VectorXd& x_min,
+                            const Eigen::VectorXd& x_max,
+                            const Eigen::VectorXd& u_min,
+                            const Eigen::VectorXd& u_max) {
+  x_min_ = x_min;
+  x_max_ = x_max;
+  u_min_ = u_min;
+  u_max_ = u_max;
+  has_constraints_ = true;
+}
+
+void ADMM::setReference(const Eigen::MatrixXd& xref) {
+  xref_ = xref;
+}
+
+void ADMM::setInitialState(const Eigen::VectorXd& x0) {
+  x0_ = x0;
+  x_solution_.col(0) = x0;
+  z_.col(0) = x0;
+}
+
+bool ADMM::solve() {
+  if (xref_.rows() != n_ || xref_.cols() != T_ + 1) {
+    return false;
+  }
+
+  // Initialize state trajectory by forward simulation
+  x_solution_.col(0) = x0_;
+  for (int t = 0; t < T_; ++t) {
+    x_solution_.col(t + 1) = A_ * x_solution_.col(t) + B_ * u_solution_.col(t);
+  }
+
+  double primal_res_x = 0.0, primal_res_u = 0.0;
+  double dual_res_x = 0.0, dual_res_u = 0.0;
+
+  for (int iter = 0; iter < max_iterations_; ++iter) {
+    z_old_ = z_;
+    u_old_ = u_;
+
+    // === X-update (state update) ===
+    // Solve: min_{x} rho/2 * ||x - z + lambda||^2 + sum (x_k - xref_k)' Q (x_k - xref_k)
+    // Subject to: x_{k+1} = A x_k + B u_k
+    for (int t = 0; t <= T_; ++t) {
+      Eigen::VectorXd x_t = z_.col(t) - lambda_x_.col(t) / rho_;
+
+      if (t == 0) {
+        // Initial state constraint
+        x_t = x0_;
+      } else if (t <= T_) {
+        // Cost gradient: 2 * Q * (x_t - xref_t)
+        Eigen::VectorXd grad = Q_.asDiagonal() * (x_t - xref_.col(t));
+        x_t = x_t - (1.0 / (rho_ + 2.0 * Q_(t % Q_.size()))) * grad;
+      }
+
+      // Apply state constraints
+      if (has_constraints_ && t > 0) {
+        x_t = x_t.cwiseMax(x_min_).cwiseMin(x_max_);
+      }
+
+      z_.col(t) = x_t;
+    }
+
+    // === U-update (control update) ===
+    // Solve: min_{u} rho/2 * ||u - v + lambda||^2 + sum u_k' R u_k
+    for (int t = 0; t < T_; ++t) {
+      // Compute the implied state from previous state and control
+      Eigen::VectorXd x_implied = A_ * x_solution_.col(t) + B_ * u_solution_.col(t);
+
+      // Gradient of cost w.r.t. u: 2 * R * u_k + B' * lambda_u
+      Eigen::VectorXd grad_u = (2.0 * R_.asDiagonal() * u_solution_.col(t)
+                                + B_.transpose() * lambda_u_.col(t)) / rho_;
+
+      Eigen::VectorXd u_t = u_solution_.col(t) - grad_u;
+
+      // Apply control constraints
+      if (has_constraints_) {
+        u_t = u_t.cwiseMax(u_min_).cwiseMin(u_max_);
+      }
+
+      u_.col(t) = u_t;
+
+      // Update state trajectory using new control
+      if (t < T_) {
+        x_solution_.col(t + 1) = A_ * x_solution_.col(t) + B_ * u_t;
+      }
+    }
+
+    // === Lagrange multiplier update ===
+    // lambda_x += rho * (x_solution - z_)
+    // lambda_u += rho * (u_solution - u_)
+    lambda_x_ += rho_ * (x_solution_ - z_);
+    lambda_u_ += rho_ * (u_solution_ - u_);
+
+    // === Check convergence ===
+    if (checkConvergence(x_solution_, u_solution_)) {
+      if (verbose_) {
+        std::cout << "ADMM converged at iteration " << iter << std::endl;
+      }
+      return true;
+    }
+
+    // === Adaptive rho adjustment (optional) ===
+    // Increase rho if primal residual is large, decrease if dual residual is large
+    // This is simplified for TinyMPC
+
+    // Update solution for next iteration
+    z_ = x_solution_;
+    u_ = u_solution_;
+  }
+
+  if (verbose_) {
+    std::cout << "ADMM reached max iterations " << max_iterations_ << std::endl;
+  }
+
+  return false;
+}
+
+bool ADMM::checkConvergence(const Eigen::MatrixXd& x, const Eigen::MatrixXd& u) {
+  // Compute primal residuals
+  double eps_pri_x = 0.0, eps_pri_u = 0.0;
+  double eps_dual_x = 0.0, eps_dual_u = 0.0;
+
+  // Primal residual for x: ||x - z||
+  for (int t = 0; t <= T_; ++t) {
+    eps_pri_x += (x.col(t) - z_.col(t)).squaredNorm();
+    eps_dual_x += (rho_ * (z_.col(t) - z_old_.col(t))).squaredNorm();
+  }
+
+  // Primal residual for u: ||u - u_old||
+  for (int t = 0; t < T_; ++t) {
+    eps_pri_u += (u.col(t) - u_.col(t)).squaredNorm();
+    eps_dual_u += (rho_ * (u_.col(t) - u_old_.col(t))).squaredNorm();
+  }
+
+  double eps_pri = std::sqrt(eps_pri_x + eps_pri_u);
+  double eps_dual = std::sqrt(eps_dual_x + eps_dual_u);
+
+  // Compute tolerance
+  double tol_pri = std::sqrt(static_cast<double>((T_ + 1) * n_ + T_ * m_)) * abs_tol_
+                   + rel_tol_ * std::max(std::sqrt(eps_pri_x), std::sqrt(eps_dual_x));
+  double tol_dual = std::sqrt(static_cast<double>((T_ + 1) * n_ + T_ * m_)) * abs_tol_
+                    + rel_tol_ * std::sqrt(eps_dual_x + eps_dual_u);
+
+  return (eps_pri < tol_pri * tol_pri) && (eps_dual < tol_dual * tol_dual);
+}
+
+}  // namespace rm_tinympc

src/rm_tinympc/src/codegen.cpp

@@ -0,0 +1,56 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "rm_tinympc/codegen.hpp"
+#include <cmath>
+
+namespace rm_tinympc {
+
+Eigen::Vector2d CodeGen::computeGimbalCommand(
+  const Eigen::Vector3d& target_pos,
+  const Eigen::Vector3d& target_vel,
+  double target_yaw,
+  double v_yaw,
+  double bullet_speed,
+  double gravity) {
+  // Simplified trajectory compensation
+  double dist = target_pos.norm();
+  double flying_time = dist / bullet_speed;
+
+  // Predict target position
+  Eigen::Vector3d predicted_pos = target_pos + flying_time * target_vel;
+
+  // Compute gimbal angles
+  double yaw = std::atan2(predicted_pos.y(), predicted_pos.x());
+  double pitch = std::atan2(predicted_pos.z(), predicted_pos.head(2).norm());
+
+  return Eigen::Vector2d(yaw, pitch);
+}
+
+Eigen::Vector3d CodeGen::predictTargetPosition(
+  const Eigen::Vector3d& pos,
+  const Eigen::Vector3d& vel,
+  double dt) {
+  return pos + dt * vel;
+}
+
+double CodeGen::computeFlyingTime(
+  const Eigen::Vector3d& target_pos,
+  double bullet_speed,
+  double gravity) {
+  double dist = target_pos.norm();
+  return dist / bullet_speed;
+}
+
+}  // namespace rm_tinympc

src/rm_tinympc/src/tiny_api.cpp

@@ -0,0 +1,71 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "rm_tinympc/tiny_api.hpp"
+
+namespace rm_tinympc {
+
+void TinyMPC::init(int n, int m, int T) {
+  n_ = n;
+  m_ = m;
+  T_ = T;
+
+  x_solution_.resize((T + 1) * n);
+  u_solution_.resize(T * m);
+  x_solution_.setZero();
+  u_solution_.setZero();
+}
+
+void TinyMPC::setProblem(
+  const Eigen::MatrixXd& A,
+  const Eigen::MatrixXd& B,
+  const Eigen::VectorXd& x0,
+  const Eigen::VectorXd& xref,
+  const Eigen::VectorXd& uref) {
+  A_ = A;
+  B_ = B;
+  x0_ = x0;
+  xref_ = xref;
+  uref_ = uref;
+}
+
+void TinyMPC::setWeights(const Eigen::VectorXd& Q, const Eigen::VectorXd& R, const Eigen::VectorXd& Qf) {
+  Q_ = Q;
+  R_ = R;
+  Qf_ = Qf;
+}
+
+void TinyMPC::solve() {
+  // Initialize with reference trajectory
+  x_solution_.segment(0, n_) = x0_;
+
+  // Simplified forward simulation using the dynamics
+  for (int t = 0; t < T_; ++t) {
+    int x_idx = t * n_;
+    int u_idx = t * m_;
+
+    // Use reference control if not set
+    if (t * m_ < uref_.size()) {
+      u_solution_.segment(u_idx, m_) = uref_.segment(u_idx, m_);
+    }
+
+    // Simulate dynamics: x_{t+1} = A * x_t + B * u_t
+    if ((t + 1) * n_ < x_solution_.size()) {
+      x_solution_.segment((t + 1) * n_, n_) = A_ * x_solution_.segment(x_idx, n_) +
+                                                B_ * u_solution_.segment(u_idx, m_);
+    }
+  }
+}
+
+}  // namespace rm_tinympc

src/rm_tinympc/src/trajectory.cpp

@@ -0,0 +1,163 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "rm_tinympc/trajectory.hpp"
+#include <cmath>
+
+namespace rm_tinympc {
+
+void QuinticPolynomial::compute(double p0, double v0, double a0,
+                                double pf, double vf, double af, double T) {
+  T_ = T;
+  double T2 = T * T;
+  double T3 = T2 * T;
+  double T4 = T3 * T;
+  double T5 = T4 * T;
+
+  // Solve for coefficients using boundary conditions
+  // p(0) = p0, p(T) = pf
+  // v(0) = v0, v(T) = vf
+  // a(0) = a0, a(T) = af
+  Eigen::Matrix3d A;
+  A << T3, T4, T5,
+       3 * T2, 4 * T3, 5 * T4,
+       6 * T, 12 * T2, 20 * T3;
+
+  Eigen::Vector3d b;
+  b << pf - p0 - v0 * T - 0.5 * a0 * T2,
+       vf - v0 - a0 * T,
+       af - a0;
+
+  Eigen::Vector3d x = A.colPivHouseholderQr().solve(b);
+
+  coef_ << p0, v0, 0.5 * a0, x[0], x[1], x[2];
+}
+
+double QuinticPolynomial::evaluate(double t) const {
+  double t2 = t * t;
+  double t3 = t2 * t;
+  double t4 = t3 * t;
+  double t5 = t4 * t;
+  return coef_[0] + coef_[1] * t + coef_[2] * t2 + coef_[3] * t3 + coef_[4] * t4 + coef_[5] * t5;
+}
+
+double QuinticPolynomial::evaluateVelocity(double t) const {
+  double t2 = t * t;
+  double t3 = t2 * t;
+  double t4 = t3 * t;
+  return coef_[1] + 2 * coef_[2] * t + 3 * coef_[3] * t2 + 4 * coef_[4] * t3 + 5 * coef_[5] * t4;
+}
+
+double QuinticPolynomial::evaluateAcceleration(double t) const {
+  double t2 = t * t;
+  double t3 = t2 * t;
+  return 2 * coef_[2] + 6 * coef_[3] * t + 12 * coef_[4] * t2 + 20 * coef_[5] * t3;
+}
+
+std::vector<TrajectoryPoint1D> TrajectoryGenerator1D::generate(
+    double p0, double pf, double max_v, double max_a, double T) {
+  std::vector<TrajectoryPoint1D> trajectory;
+
+  // Compute quintic polynomial
+  poly_.compute(p0, 0, 0, pf, 0, 0, T);
+
+  // Sample trajectory
+  const int num_samples = static_cast<int>(T * 100);  // 100 Hz
+  for (int i = 0; i <= num_samples; ++i) {
+    double t = static_cast<double>(i) / num_samples * T;
+    TrajectoryPoint1D pt;
+    pt.t = t;
+    pt.p = poly_.evaluate(t);
+    pt.v = poly_.evaluateVelocity(t);
+    pt.a = poly_.evaluateAcceleration(t);
+    trajectory.push_back(pt);
+  }
+
+  return trajectory;
+}
+
+double TrajectoryGenerator1D::generateMinTime(
+    double p0, double pf, double max_v, double max_a) {
+  double d = std::abs(pf - p0);
+
+  // Minimum time for bang-bang acceleration profile
+  // t_acc = max_v / max_a
+  // d_acc = 0.5 * max_a * t_acc^2 = max_v^2 / (2 * max_a)
+  // If distance < 2 * d_acc, we're limited by acceleration
+  // If distance >= 2 * d_acc, we can reach max_v
+
+  double t_acc = max_v / max_a;
+  double d_acc = 0.5 * max_a * t_acc * t_acc;  // Distance for acceleration
+  double d_total = 2 * d_acc;  // Distance for accel + decel
+
+  if (d <= d_total) {
+    // Triangle profile: can't reach max_v
+    // d = 0.5 * max_a * t^2 for triangle
+    T_ = std::sqrt(4 * d / max_a);
+  } else {
+    // Trapezoidal profile: reach max_v
+    double d_cruise = d - d_total;
+    double t_cruise = d_cruise / max_v;
+    T_ = 2 * t_acc + t_cruise;
+  }
+
+  return T_;
+}
+
+std::vector<TrajectoryPoint2D> GimbalTrajectoryGenerator::generate(
+    double yaw0, double pitch0, double yawf, double pitchf,
+    double max_yaw_rate, double max_pitch_rate,
+    double max_yaw_acc, double max_pitch_acc, double T) {
+
+  std::vector<TrajectoryPoint2D> trajectory;
+
+  // Compute quintic polynomials for yaw and pitch
+  poly_yaw_.compute(yaw0, 0, 0, yawf, 0, 0, T);
+  poly_pitch_.compute(pitch0, 0, 0, pitchf, 0, 0, T);
+
+  // Sample trajectory
+  const int num_samples = static_cast<int>(T * 100);  // 100 Hz
+  for (int i = 0; i <= num_samples; ++i) {
+    double t = static_cast<double>(i) / num_samples * T;
+    TrajectoryPoint2D pt;
+    pt.t = t;
+    pt.yaw = poly_yaw_.evaluate(t);
+    pt.yaw_vel = poly_yaw_.evaluateVelocity(t);
+    pt.yaw_acc = poly_yaw_.evaluateAcceleration(t);
+    pt.pitch = poly_pitch_.evaluate(t);
+    pt.pitch_vel = poly_pitch_.evaluateVelocity(t);
+    pt.pitch_acc = poly_pitch_.evaluateAcceleration(t);
+    trajectory.push_back(pt);
+  }
+
+  return trajectory;
+}
+
+double GimbalTrajectoryGenerator::generateMinTime(
+    double yaw0, double pitch0, double yawf, double pitchf,
+    double max_yaw_rate, double max_pitch_rate,
+    double max_yaw_acc, double max_pitch_acc) {
+
+  TrajectoryGenerator1D yaw_gen, pitch_gen;
+
+  double T_yaw = yaw_gen.generateMinTime(yaw0, yawf, max_yaw_rate, max_yaw_acc);
+  double T_pitch = pitch_gen.generateMinTime(pitch0, pitchf, max_pitch_rate, max_pitch_acc);
+
+  // Use the maximum of the two times to ensure both constraints are satisfied
+  T_ = std::max(T_yaw, T_pitch);
+
+  return T_;
+}
+
+}  // namespace rm_tinympc

src/rm_utils/include/rm_utils/math/error_state_kalman_filter.hpp

@@ -0,0 +1,179 @@
+// Copyright (C) FYT Vision Group. All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef ERROR_STATE_KALMAN_FILTER_HPP_
+#define ERROR_STATE_KALMAN_FILTER_HPP_
+
+#include <Eigen/Eigen>
+#include <functional>
+
+namespace fyt {
+
+template <int X_N, int Z_N, typename PredictModel, typename MeasureModel>
+class ErrorStateKalmanFilter {
+public:
+  using StateVector = Eigen::Vector<double, X_N>;
+  using MeasurementVector = Eigen::Vector<double, Z_N>;
+  using StateMatrix = Eigen::Matrix<double, X_N, X_N>;
+  using MeasurementMatrix = Eigen::Matrix<double, Z_N, Z_N>;
+  using CrossCovarianceMatrix = Eigen::Matrix<double, X_N, Z_N>;
+  using JacobianMatrix = Eigen::Matrix<double, X_N, X_N>;
+
+  ErrorStateKalmanFilter() = default;
+
+  void init(const StateVector& initial_state,
+            const StateMatrix& initial_covariance,
+            const MeasurementMatrix& measurement_noise,
+            const StateMatrix& process_noise) {
+    state_ = initial_state;
+    covariance_ = initial_covariance;
+    R_ = measurement_noise;
+    Q_ = process_noise;
+  }
+
+  void setNoiseMatrices(const MeasurementMatrix& R, const StateMatrix& Q) {
+    R_ = R;
+    Q_ = Q;
+  }
+
+  void setResidualFunc(std::function<MeasurementVector(const MeasurementVector&,
+                                                       const MeasurementVector&)> func) {
+    residual_func_ = func;
+  }
+
+  void setInjectFunc(std::function<void(StateVector&, const StateVector&)> func) {
+    inject_func_ = func;
+  }
+
+  StateVector predict(double dt) {
+    PredictModel predict_model(dt);
+
+    // State prediction using the motion model
+    StateVector predicted_state = state_;
+    predict_model(state_.data(), predicted_state.data());
+
+    // Jacobian of the motion model w.r.t. state
+    JacobianMatrix F = JacobianMatrix::Identity();
+    // This is a simplified Jacobian - for complex models, numerical differentiation
+    // or analytical Jacobians should be used
+    // For now, we use identity as the motion is linear in velocity terms
+
+    // Covariance prediction
+    covariance_ = F * covariance_ * F.transpose() + Q_;
+
+    state_ = predicted_state;
+    return state_;
+  }
+
+  StateVector update(const MeasurementVector& measurement) {
+    MeasurementVector z_predicted;
+    MeasureModel measure_model;
+    measure_model(state_.data(), z_predicted.data());
+
+    // Innovation (measurement residual)
+    MeasurementVector y;
+    if (residual_func_) {
+      y = residual_func_(measurement, z_predicted);
+    } else {
+      y = measurement - z_predicted;
+    }
+
+    // Innovation covariance
+    MeasurementMatrix S = measurement_jacobian_ * covariance_ * measurement_jacobian_.transpose() + R_;
+
+    // Kalman gain
+    CrossCovarianceMatrix Pxz = covariance_ * measurement_jacobian_.transpose();
+    Eigen::Matrix<double, X_N, Z_N> K = Pxz * S.inverse();
+
+    // State update
+    StateVector delta_x = K * y;
+    state_ = state_ + delta_x;
+
+    // Covariance update (Joseph form for numerical stability)
+    Eigen::Matrix<double, X_N, X_N> I_KH = JacobianMatrix::Identity() - K * measurement_jacobian_;
+    covariance_ = I_KH * covariance_ * I_KH.transpose() + K * R_ * K.transpose();
+
+    return state_;
+  }
+
+  StateVector updateIterative(const MeasurementVector& measurement, int max_iterations = 10) {
+    StateVector updated_state = state_;
+    MeasurementMatrix H = measurement_jacobian_;
+
+    for (int i = 0; i < max_iterations; ++i) {
+      MeasurementVector z_predicted;
+      MeasureModel measure_model;
+      measure_model(updated_state.data(), z_predicted.data());
+
+      MeasurementVector y;
+      if (residual_func_) {
+        y = residual_func_(measurement, z_predicted);
+      } else {
+        y = measurement - z_predicted;
+      }
+
+      MeasurementMatrix S = H * covariance_ * H.transpose() + R_;
+      Eigen::Matrix<double, X_N, Z_N> K = covariance_ * H.transpose() * S.inverse();
+
+      StateVector delta_x = K * y;
+      updated_state = updated_state + delta_x;
+
+      if (delta_x.norm() < 1e-6) {
+        break;
+      }
+    }
+
+    if (inject_func_) {
+      inject_func_(state_, updated_state - state_);
+    }
+    state_ = updated_state;
+
+    return state_;
+  }
+
+  void setState(const StateVector& state) {
+    state_ = state;
+  }
+
+  StateVector getState() const {
+    return state_;
+  }
+
+  StateMatrix getCovariance() const {
+    return covariance_;
+  }
+
+  double getNIS() const {
+    return last_nis_;
+  }
+
+  void setMeasurementJacobian(const MeasurementMatrix& H) {
+    measurement_jacobian_ = H;
+  }
+
+private:
+  StateVector state_;
+  StateMatrix covariance_;
+  MeasurementMatrix R_;  // Measurement noise covariance
+  StateMatrix Q_;        // Process noise covariance
+  MeasurementMatrix measurement_jacobian_;
+  double last_nis_ = 0;
+
+  std::function<MeasurementVector(const MeasurementVector&, const MeasurementVector&)> residual_func_;
+  std::function<void(StateVector&, const StateVector&)> inject_func_;
+};
+
+}  // namespace fyt
+
+#endif  // ERROR_STATE_KALMAN_FILTER_HPP_

src/rm_utils/include/rm_utils/math/extended_kalman_filter.hpp

@@ -48,6 +48,8 @@ public:
 
   using UpdateQFunc = std::function<MatrixXX()>;
   using UpdateRFunc = std::function<MatrixZZ(const MatrixZ1 &z)>;
+  // Residual function for handling angle wraparound (e.g., -pi~pi)
+  using ResidualFunc = std::function<MatrixZ1(const MatrixZ1 &z_meas, const MatrixZ1 &z_pred)>;
 
   explicit ExtendedKalmanFilter(const PredicFunc &f,
                                 const MeasureFunc &h,
@@ -66,6 +68,14 @@ public:
 
   void setMeasureFunc(const MeasureFunc &h) noexcept { this->h = h; }
 
+  // Set residual function for handling angle wraparound
+  void setResidualFunc(const ResidualFunc &residual_func) noexcept {
+    residual_func_ = residual_func;
+  }
+
+  // Check if residual function is set
+  bool hasResidualFunc() const noexcept { return residual_func_ != nullptr; }
+
   // Compute a predicted state
   MatrixX1 predict() noexcept {
     ceres::Jet<double, N_X> x_e_jet[N_X];
@@ -106,14 +116,66 @@ public:
     }
 
     R = update_R(z);
+
+    // Compute innovation (measurement residual)
+    // Use residual_func if set to handle angle wraparound
+    if (residual_func_ != nullptr) {
+      innovation_ = residual_func_(z, z_pri_);
+    } else {
+      innovation_ = z - z_pri_;
+    }
+
     S_ = H * P_pri * H.transpose() + R;
     K = P_pri * H.transpose() * S_.inverse();
-    innovation_ = z - z_pri_;
+
     x_post = x_post + K * innovation_;
     P_post = (MatrixXX::Identity() - K * H) * P_pri;
     return x_post;
   }
 
+  // Update with iterative measurement update for better angle handling
+  MatrixX1 updateIterative(const MatrixZ1 &z, int max_iterations = 3) noexcept {
+    MatrixX1 updated_state = x_post;
+
+    for (int iter = 0; iter < max_iterations; ++iter) {
+      ceres::Jet<double, N_X> x_jet[N_X];
+      for (int i = 0; i < N_X; i++) {
+        x_jet[i].a = updated_state[i];
+        x_jet[i].v[i] = 1;
+      }
+      ceres::Jet<double, N_X> z_jet[N_Z];
+      h(x_jet, z_jet);
+
+      MatrixZ1 z_pred;
+      for (int i = 0; i < N_Z; i++) {
+        z_pred[i] = z_jet[i].a;
+      }
+
+      R = update_R(z);
+
+      MatrixZ1 innovation;
+      if (residual_func_ != nullptr) {
+        innovation = residual_func_(z, z_pred);
+      } else {
+        innovation = z - z_pred;
+      }
+
+      MatrixZZ S = H * P_pri * H.transpose() + R;
+      MatrixXZ K = P_pri * H.transpose() * S.inverse();
+
+      MatrixX1 delta_x = K * innovation;
+      updated_state = updated_state + delta_x;
+
+      // Check convergence
+      if (delta_x.norm() < 1e-6) {
+        break;
+      }
+    }
+
+    x_post = updated_state;
+    return x_post;
+  }
+
   // Get innovation (z - z_pri)
   const MatrixZ1 & getInnovation() const { return innovation_; }
 
@@ -126,6 +188,15 @@ public:
     return innovation_.transpose() * S_.inverse() * innovation_;
   }
 
+  // Get Kalman gain
+  const MatrixXZ & getKalmanGain() const { return K; }
+
+  // Get posterior state
+  const MatrixX1 & getState() const { return x_post; }
+
+  // Get posterior covariance
+  const MatrixXX & getCovariance() const { return P_post; }
+
 private:
   // Process nonlinear vector function
   PredicFunc f;
@@ -139,6 +210,8 @@ private:
   // Measurement noise covariance matrix
   UpdateRFunc update_R;
   MatrixZZ R;
+  // Residual function for angle wraparound handling
+  ResidualFunc residual_func_;
 
   // Priori error estimate covariance matrix
   MatrixXX P_pri;