v0.1.8; add configs, fix bugs

UPDATE.md

@@ -1,8 +1,16 @@
+# 20260109
+## v0.1.8
+1. 加入两个配置legged_gym/envs/go2/go2_config_vanilla.py和legged_gym/envs/go2/go2_config_vanilla2.py
+    - vanilla: episode长度25, 指令采样步数5, 打开move_down_by_accumulated_xy_command, dynamic_resample_commands, 关闭heading_command, zero_command_curriculum, limit_vel_prob, command_range_curriculum, dynamic_sigma
+    - vanilla2：episode长度20, 指令采样步数10, 打开haeding_command, 关闭move_down_by_accumulated_xy_command, dynamic_resample_commands, zero_command_curriculum, limit_vel_prob, command_range_curriculum, dynamic_sigma
+2. 加入go2_config_fast_flat_move.py配置
+3. onnx支持rem-cts导出
 # 20260107
 ## v0.1.7
 Fix bug: 
 1. 修复`cfg.commands.dynamic_resample_commands=False`时, 未根据环境调整指令采样范围的问题
 2. 修复`env_command_ranges`初始值问题
+3. 修复`dynamic_resample_commands=False`时, 未更新`commands_resampling_step`问题
 ## v0.1.6
 1. 加入`go2_rem_cts`, student使用MoE结构, teacher使用普通CTS, 使用非共享权重和全goal输入
 2. 加入`move_down_by_accumulated_xy_command`选择是否通过累计速度来降低等级

legged_gym/envs/base/legged_robot.py

@@ -426,6 +426,8 @@ class LeggedRobot(BaseTask):
         Args:
             env_ids (List[int]): Environments ids for which new commands are needed
         """
+        if len(env_ids) == 0:
+            return
         self.stop_heading[env_ids] = False
         # update command curriculum with train steps
         if len(self.cfg.commands.command_range_curriculum):
@@ -443,6 +445,7 @@ class LeggedRobot(BaseTask):
                     self._update_env_command_ranges()
                     print(f"Command range updated at iter {current_iter}: {self.command_ranges}")
         remaining_dist = torch.clip(0.625 * self.cfg.terrain.terrain_length - torch.norm(self.commands_xy_accumulation[env_ids], dim=1) * self.cfg.commands.resampling_time, 0.0)
+        self.commands_resampling_step[env_ids] = self.cfg.commands.resampling_time / self.dt
         if self.cfg.commands.dynamic_resample_commands:
             # arrive at boundary 0.625 times the width of the remaining distance
             if ((self.max_episode_length - self.episode_length_buf[env_ids]) == 0).any():
@@ -472,7 +475,6 @@ class LeggedRobot(BaseTask):
                 lower = self.env_command_ranges["ang_vel_yaw"][env_ids, 0]
                 upper = self.env_command_ranges["ang_vel_yaw"][env_ids, 1]
                 self.commands[env_ids, 2] = (upper - lower) * r + lower
-            self.commands_resampling_step[env_ids] = self.cfg.commands.resampling_time / self.dt
         else:
             self.commands[env_ids, 0] = sample_single_interval(
                 env_ids,
@@ -859,6 +861,12 @@ class LeggedRobot(BaseTask):
     def _update_env_command_ranges(self):
         """ Update environment-wise command ranges based on current command ranges and terrain type """
         if not hasattr(self, 'terrain_ids'):
+            self.env_command_ranges = {
+                'lin_vel_x': torch.tensor(self.command_ranges['lin_vel_x'], device=self.device, requires_grad=False).repeat(self.num_envs, 1),
+                'lin_vel_y': torch.tensor(self.command_ranges['lin_vel_y'], device=self.device, requires_grad=False).repeat(self.num_envs, 1),
+                'ang_vel_yaw': torch.tensor(self.command_ranges['ang_vel_yaw'], device=self.device, requires_grad=False).repeat(self.num_envs, 1),
+                'heading': torch.tensor(self.command_ranges['heading'], device=self.device, requires_grad=False).repeat(self.num_envs, 1),
+            }
             return
         for terrain_id, terrain_command_ranges in enumerate(self.cfg.commands.terrain_max_command_ranges):
             env_ids = (self.terrain_ids == terrain_id).nonzero(as_tuple=False).flatten()
@@ -1322,6 +1330,7 @@ class LeggedRobot(BaseTask):
             sigma_y = self._get_dynamic_sigma(torch.abs(self.commands[:, 1]), vmin, vmax)
         lin_vel_error_sq = torch.square(self.commands[:, :2] - self.base_lin_vel[:, :2])
         scaled_error = lin_vel_error_sq[:, 0] / sigma_x + lin_vel_error_sq[:, 1] / sigma_y
+        # print(f"{self.base_lin_vel[:, :2]=}, {lin_vel_error_sq=}")
         return torch.exp(-scaled_error)
     
     def _reward_tracking_ang_vel(self):

legged_gym/envs/go2/go2_config_fast_flat_move.py

@@ -0,0 +1,339 @@
+# -*- coding: utf-8 -*-
+'''
+@File    : go2_config_fast_flat_move.py
+@Time    : 2026/01/10 02:31:14
+@Author  : wty-yy
+@Version : 1.0
+@Blog    : https://wty-yy.github.io/
+@Desc    : go2 fast flat move config file
+Change command_range_curriculum, init command range
+'''
+import math
+from legged_gym.envs.base.legged_robot_config import LeggedRobotCfg, LeggedRobotCfgPPO, LeggedRobotCfgCTS, LeggedRobotCfgMoECTS, LeggedRobotCfgMoECTS, LeggedRobotCfgMCPCTS, LeggedRobotCfgACMoECTS, LeggedRobotCfgDualMoECTS, LeggedRobotCfgREMCTS
+
+class GO2Cfg(LeggedRobotCfg):
+    class init_state(LeggedRobotCfg.init_state):
+        pos = [0.0, 0.0, 0.42] # x,y,z [m]
+        default_joint_angles = { # = target angles [rad] when action = 0.0
+            'FL_hip_joint': 0.1,   # [rad]
+            'RL_hip_joint': 0.1,   # [rad]
+            'FR_hip_joint': -0.1 ,  # [rad]
+            'RR_hip_joint': -0.1,   # [rad]
+
+            'FL_thigh_joint': 0.8,     # [rad]
+            'RL_thigh_joint': 1.,   # [rad]
+            'FR_thigh_joint': 0.8,     # [rad]
+            'RR_thigh_joint': 1.,   # [rad]
+
+            'FL_calf_joint': -1.5,   # [rad]
+            'RL_calf_joint': -1.5,    # [rad]
+            'FR_calf_joint': -1.5,  # [rad]
+            'RR_calf_joint': -1.5,    # [rad]
+        }
+        turn_over = False # initialize the robot in a flipped over position
+        # turn_over_proportions = [0.1, 0.3, 0.6] # proportions for backflip, sideflip, noflip
+        turn_over_proportions = [0.0, 0.2, 0.8] # proportions for backflip, sideflip, noflip
+        turn_over_init_heights = { # initial heights range for each flip type
+            'backflip': [0.10, 0.15],
+            'sideflip': [0.16, 0.21],
+        }
+        # turn_over_proportions = [0.0, 1.0, 0.0] # proportions for backflip, sideflip, noflip
+
+    class env(LeggedRobotCfg.env):
+        num_envs = 8192
+        num_observations = 45
+        # obs(45) + base_lin_vel(3) + height_measurements(187)
+        num_privileged_obs = 45 + 3 + 4 + 12 + 12 + 187  # 263
+        # num_privileged_obs = 45 + 3 + 187  # 235
+        # num_privileged_obs = 48  # without height measurements
+        episode_length_s = 25
+
+    class domain_rand(LeggedRobotCfg.domain_rand):
+        ### Robot properties ###
+        randomize_friction = True
+        friction_range = [0.0, 2.0]
+
+        randomize_base_mass = True
+        added_mass_range = [-1., 1.]
+
+        randomize_link_mass = True
+        multiplied_link_mass_range = [0.9, 1.1]
+
+        randomize_base_com = True
+        added_base_com_range = [-0.03, 0.03]
+
+        randomize_restitution = True # restitution to robot links (Robot init)
+        restitution_range = [0.0, 0.5]
+
+        ### Environment reset ###
+        randomize_pd_gains = True
+        stiffness_multiplier_range = [0.9, 1.1]  
+        damping_multiplier_range = [0.9, 1.1]    
+
+        randomize_motor_zero_offset = True
+        motor_zero_offset_range = [-0.035, 0.035]
+
+        randomize_motor_strength = True # (Env reset)
+        motor_strength_range = [0.8, 1.2]
+
+        ### Environment step ###
+        push_robots = True
+        push_interval_s = 4
+        max_push_vel_xy = 0.4
+        max_push_ang_vel = 0.6
+
+        randomize_action_delay = True # use last_action with 0~20 ms delay, 4 decimation
+
+    class control(LeggedRobotCfg.control):
+        # PD Drive parameters:
+        control_type = 'P'
+        stiffness = {'joint': 20.0}  # [N*m/rad]
+        damping = {'joint': 0.5}     # [N*m*s/rad]
+        # action scale: target angle = actionScale * action + defaultAngle
+        action_scale = 0.25
+        # decimation: Number of control action updates @ sim DT per policy DT
+        decimation = 4
+    
+    class terrain(LeggedRobotCfg.terrain):
+        mesh_type = 'plane'
+        max_init_terrain_level = 5
+        # [wave, slope, rough_slope, stairs up, stairs down, obstacles, stepping_stones, gap, flat]
+        # terrain_proportions = [0.2, 0.05, 0.05, 0.30, 0.05, 0.25, 0.0, 0.0, 0.1]  # 更偏向wave
+        terrain_proportions = [0.05, 0.20, 0.05, 0.25, 0.10, 0.20, 0.0, 0.0, 0.15]  # 这个更偏向平地斜坡
+        # terrain_proportions = [0.20, 0.05, 0.05, 0.30, 0.15, 0.20, 0.0, 0.0, 0.05]  # 更偏向wave和stairs
+        # terrain_proportions = [0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0]
+        # terrain_proportions = [0.3, 0.3, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1]
+        # terrain_proportions = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0]
+        move_down_by_accumulated_xy_command = True # move down the terrain curriculum based on accumulated xy command distance instead of absolute distance
+        
+    class commands(LeggedRobotCfg.commands):
+        curriculum = False
+        max_curriculum = 1.
+        num_commands = 4 # default: lin_vel_x, lin_vel_y, ang_vel_yaw (in heading mode ang_vel_yaw is recomputed from heading error)
+        resampling_time = 5. # time before command are changed[s]
+        heading_command = False # if true: compute ang vel command from heading error
+        # start training with zero commands and then gradually increase zero command probability
+        zero_command_curriculum = {'start_iter': 0, 'end_iter': 1500, 'start_value': 0.0, 'end_value': 0.1}
+        limit_ang_vel_at_zero_command_prob = 0.2 # probability of add limiting angular velocity commands when zero command is sampled
+        limit_vel_prob = 0.2 # probability of limiting linear velocity command
+        limit_vel_invert_when_continuous = True # invert the limit logic when using continuous sample limit velocity commands
+        limit_vel = {"lin_vel_x": [-1, 1], "lin_vel_y": [-1, 1], "ang_vel_yaw": [-1, 0, 1]} # sample vel commands from min [-1] or zero [0] or max [1] range only
+        stop_heading_at_limit = True # stop heading updates when vel is limited
+        dynamic_resample_commands = True # sample commands with low bounds
+        command_range_curriculum = [ { # list for command range curriculums at specific training iterations
+            'iter': 5000, # training iteration at which the command ranges are updated
+            'lin_vel_x': [-2.0, 2.0], # min max [m/s]
+            'lin_vel_y': [-1.0, 1.0], # min max [m/s]
+            'ang_vel_yaw': [-2.0, 2.0], # min max [rad/s]
+            'heading': [-1.57, 1.57], # min max [rad]
+        }, { # list for command range curriculums at specific training iterations
+            'iter': 10000, # training iteration at which the command ranges are updated
+            'lin_vel_x': [-3.0, 3.0], # min max [m/s]
+            'lin_vel_y': [-1.0, 1.0], # min max [m/s]
+            'ang_vel_yaw': [-2.0, 2.0], # min max [rad/s]
+            'heading': [-1.57, 1.57], # min max [rad]
+        }, { # list for command range curriculums at specific training iterations
+            'iter': 20000, # training iteration at which the command ranges are updated
+            'lin_vel_x': [-3.5, 3.5], # min max [m/s]
+            'lin_vel_y': [-0.5, 0.5], # min max [m/s]
+            'ang_vel_yaw': [-1.0, 1.0], # min max [rad/s]
+            'heading': [-1.57, 1.57], # min max [rad]
+        }, { # list for command range curriculums at specific training iterations
+            'iter': 30000, # training iteration at which the command ranges are updated
+            'lin_vel_x': [-4.0, 4.0], # min max [m/s]
+            'lin_vel_y': [-0.0, 0.0], # min max [m/s]
+            'ang_vel_yaw': [-0.0, 0.0], # min max [rad/s]
+            'heading': [-1.57, 1.57], # min max [rad]
+        }, { # list for command range curriculums at specific training iterations
+            'iter': 40000, # training iteration at which the command ranges are updated
+            'lin_vel_x': [-4.5, 4.5], # min max [m/s]
+            'lin_vel_y': [-0.0, 0.0], # min max [m/s]
+            'ang_vel_yaw': [-0.0, 0.0], # min max [rad/s]
+            'heading': [-1.57, 1.57], # min max [rad]
+        }
+        ]
+        turn_over_zero_time = { # if turn_over is true, time robot must be stable before sampling new commands after a turn over
+            "backflip": 5.0,
+            "sideflip": 3.0,
+        }
+        # [wave, slope, rough slope, stairs up, stairs down, obstacles, stepping stones, gap, flat]
+        terrain_max_command_ranges = [
+            {'lin_vel_x': [-1.5, 1.5], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # wave
+            {'lin_vel_x': [-1.5, 1.5], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # slope
+            {'lin_vel_x': [-1.5, 1.5], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # rough slope
+            {'lin_vel_x': [-1.0, 1.0], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # stairs up
+            {'lin_vel_x': [-1.0, 1.0], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # stairs down
+            {'lin_vel_x': [-1.0, 1.0], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # obstacles
+            {'lin_vel_x': [-1.0, 1.0], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # stepping stones
+            {'lin_vel_x': [-1.0, 1.0], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # gap
+            {'lin_vel_x': [-2.0, 2.0], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-2.0, 2.0], 'heading': [-1.57, 1.57]},  # flat
+        ]
+
+        class ranges:
+            lin_vel_x = [-1.0, 1.0] # min max [m/s]
+            lin_vel_y = [-1.0, 1.0] # min max [m/s]
+            ang_vel_yaw = [-1.0, 1.0]   # min max [rad/s]
+            heading = [-1.57, 1.57] # min max [rad]
+        
+    class asset(LeggedRobotCfg.asset):
+        file = '{LEGGED_GYM_ROOT_DIR}/resources/robots/go2/urdf/go2.urdf'
+        name = "go2"
+        foot_name = "foot"
+        penalize_contacts_on = ["thigh", "calf"]
+        terminate_after_contacts_on = ["base"]
+        self_collisions = 1 # 1 to disable, 0 to enable...bitwise filter
+  
+    class rewards(LeggedRobotCfg.rewards):
+        soft_dof_pos_limit = 0.9
+        base_height_target = 0.38
+        only_positive_rewards = False
+        max_contact_force = 147. # forces above this value are penalized, go2 weight 15kg
+        curriculum_rewards = [
+            {'reward_name': 'lin_vel_z', 'start_iter': 0, 'end_iter': 1500, 'start_value': 1.0, 'end_value': 0.0},
+            {'reward_name': 'correct_base_height', 'start_iter': 0, 'end_iter': 5000, 'start_value': 1.0, 'end_value': 10.0},
+            # {'reward_name': 'dof_power', 'start_iter': 0, 'end_iter': 3000, 'start_value': 1.0, 'end_value': 0.1},
+            # {'reward_name': 'upright', 'start_iter': 0, 'end_iter': 1500, 'start_value': 1.0, 'end_value': 0.0},
+        ]
+        tracking_sigma = 0.25 # tracking reward = exp(-error^2/sigma)
+        dynamic_sigma = { # linear interpolation of sigma based on command velocity, **Must start terrain curriculum first**
+            "min_lin_vel": 0.5, # min abs linear velocity to have default sigma
+            "max_lin_vel": 1.5, # max abs linear velocity to have max sigma
+            "min_ang_vel": 1.0, # min abs angular velocity to have default sigma
+            "max_ang_vel": 2.0, # max abs angular velocity to have max sigma
+            # wave, slope, rough_slope, stairs up, stairs down, obstacles, stepping_stones, gap, flat]
+            # "max_sigma": [1/3, 1/4, 1/4, 1/2.7, 1/2.7, 1/2, 1, 1, 1/4]
+            "max_sigma": [5/12, 1/4, 1/4, 1/2, 1/2, 3/4, 1, 1, 1/4]
+        }
+        min_legs_distance = 0.1  # min distance between legs to not be considered stumbling
+        class scales:
+            # tracking_lin_vel = 1.0
+            # tracking_ang_vel = 0.2
+            # lin_vel_z = -10.0
+            # base_height = -50.0
+            # action_rate = -0.005
+            # similar_to_default = -0.1
+            # dof_power = -1e-3  # 能够明显抑制跳跃
+            # dof_acc = -3e-7
+
+            # tracking_lin_vel = 1.0
+            # tracking_ang_vel = 0.5
+            # lin_vel_z = -2.0
+            # ang_vel_xy = -0.05
+            # dof_acc = -2.5e-7
+            # dof_power = -1e-3  # 能够明显抑制跳跃
+            # # torques = -1e-4  # 无用会走着走着倒了
+            # correct_base_height = -10.0
+            # action_rate = -0.01
+            # action_smoothness = -0.01
+            # collision = -1.0
+            # dof_pos_limits = -2.0
+            # feet_regulation = -0.05
+            # hip_to_default = -0.1
+            # similar_to_default = -0.05
+
+            # CTS reward
+            tracking_lin_vel = 1.0
+            tracking_ang_vel = 0.5
+            lin_vel_z = -2.0
+            ang_vel_xy = -0.05
+            dof_acc = -2.5e-7
+            dof_power = -2e-5
+            torques = -1e-4
+            correct_base_height = -1.0
+            action_rate = -0.01
+            action_smoothness = -0.01
+            collision = -1.0
+            dof_pos_limits = -2.0
+            feet_regulation = -0.05
+            # CTS奖励训出来双脚距离非常近, 真机效果很差, 但是sim2sim能上20cm楼梯, 尝试加入hip_to_default奖励或similar_to_default奖励
+            hip_to_default = -0.05  # 在训练到y=1.5时, 双脚会明显碰撞, 为避免该问题提升hip, 效果更差, 还是保持0.05 (y最大也只到0.1了)
+            # legs_distance = -1.5  # 奖励双脚距离, 避免CTS训练出来双脚距离过近, 尝试加入后robogauge flat验证效果变差, 删除
+            # similar_to_default = -0.01
+            # feet_contact_forces = -1.0  # 尝试加入但并没有起到任何效果, 删除
+
+        turn_over_roll_threshold = math.pi / 4 # threshold on roll to use turn over rewards
+        class turn_over_scales:
+            upright = 1.0
+            # dof_acc = -2.5e-7
+            # dof_power = -2e-5
+            # action_rate = -0.001
+            # action_smoothness = -0.001
+
+    class noise(LeggedRobotCfg.noise):
+        add_noise = True
+
+class GO2CfgPPO(LeggedRobotCfgPPO):
+    class algorithm(LeggedRobotCfgPPO.algorithm):
+        entropy_coef = 0.01
+    class runner(LeggedRobotCfgPPO.runner):
+        run_name = ''
+        experiment_name = 'go2_ppo'
+        max_iterations = 100000
+        save_interval = 500
+
+class GO2CfgCTS(LeggedRobotCfgCTS):
+    class runner(LeggedRobotCfgCTS.runner):
+        num_steps_per_env = 24
+        run_name = ''
+        experiment_name = 'go2_cts'
+        max_iterations = 150000
+        save_interval = 500
+    
+    class policy(LeggedRobotCfgCTS.policy):
+        latent_dim = 32
+        norm_type = 'l2norm'
+
+class GO2CfgMoECTS(LeggedRobotCfgMoECTS):
+    class policy(LeggedRobotCfgMoECTS.policy):
+        obs_no_goal_mask = [True] * 6 + [False] * 3 + [True] * 36  # mask for obs without command info
+        student_expert_num = 8 # number of experts in the student model
+    
+    class algorithm(LeggedRobotCfgMoECTS.algorithm):
+        load_balance_coef = 0.01
+
+    class runner(LeggedRobotCfgMoECTS.runner):
+        run_name = ''
+        experiment_name = 'go2_moe_cts'
+        max_iterations = 150000
+        save_interval = 500
+
+class GO2CfgMCPCTS(LeggedRobotCfgMCPCTS):
+    class policy(LeggedRobotCfgMCPCTS.policy):
+        obs_no_goal_mask = [True] * 6 + [False] * 3 + [True] * 36  # mask for obs without command info
+        student_expert_num = 8 # number of experts in the student model
+
+    class runner(LeggedRobotCfgMCPCTS.runner):
+        run_name = ''
+        experiment_name = 'go2_mcp_cts'
+        max_iterations = 150000
+        save_interval = 500
+
+class GO2CfgACMoECTS(LeggedRobotCfgACMoECTS):
+    class policy(LeggedRobotCfgACMoECTS.policy):
+        expert_num = 8  # number of experts in the student model
+    
+    class runner(LeggedRobotCfgACMoECTS.runner):
+        run_name = ''
+        experiment_name = 'go2_ac_moe_cts'
+        max_iterations = 150000
+        save_interval = 500
+
+class GO2CfgDualMoECTS(LeggedRobotCfgDualMoECTS):
+    class policy(LeggedRobotCfgDualMoECTS.policy):
+        expert_num = 8  # number of experts in the student model
+    
+    class runner(LeggedRobotCfgDualMoECTS.runner):
+        run_name = ''
+        experiment_name = 'go2_dual_moe_cts'
+        max_iterations = 150000
+        save_interval = 500
+
+class GO2CfgREMCTS(LeggedRobotCfgREMCTS):
+    class policy(LeggedRobotCfgREMCTS.policy):
+        expert_num = 8  # number of experts in the student model
+    
+    class runner(LeggedRobotCfgREMCTS.runner):
+        run_name = ''
+        experiment_name = 'go2_rem_cts'
+        max_iterations = 150000
+        save_interval = 500

legged_gym/envs/go2/go2_config_vanilla.py

@@ -1,4 +1,15 @@
-# Don't forget to change IS_HARD = False, if you want to use original training setting
+# -*- coding: utf-8 -*-
+'''
+@File    : go2_config_vanilla.py
+@Time    : 2026/01/10 02:26:04
+@Author  : wty-yy
+@Version : 1.0
+@Blog    : https://wty-yy.github.io/
+@Desc    : Go2 vanilla training config
+episode length 25, resample commands 5 sec,
+open move_down_by_accumulated_xy_command, dynamic_resample_commands
+close heading_command, zero_command_curriculum, limit_vel_prob, command_range_curriculum, dynamic_sigma
+'''
 import math
 from legged_gym.envs.base.legged_robot_config import LeggedRobotCfg, LeggedRobotCfgPPO, LeggedRobotCfgCTS, LeggedRobotCfgMoECTS, LeggedRobotCfgMoECTS, LeggedRobotCfgMCPCTS, LeggedRobotCfgACMoECTS, LeggedRobotCfgDualMoECTS, LeggedRobotCfgREMCTS
 
@@ -94,7 +105,7 @@ class GO2Cfg(LeggedRobotCfg):
         # terrain_proportions = [0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0]
         # terrain_proportions = [0.3, 0.3, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1]
         # terrain_proportions = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0]
-        move_down_by_accumulated_xy_command = False # move down the terrain curriculum based on accumulated xy command distance instead of absolute distance
+        move_down_by_accumulated_xy_command = True # move down the terrain curriculum based on accumulated xy command distance instead of absolute distance
         
     class commands(LeggedRobotCfg.commands):
         curriculum = False
@@ -110,7 +121,7 @@ class GO2Cfg(LeggedRobotCfg):
         limit_vel_invert_when_continuous = True # invert the limit logic when using continuous sample limit velocity commands
         limit_vel = {"lin_vel_x": [-1, 1], "lin_vel_y": [-1, 1], "ang_vel_yaw": [-1, 0, 1]} # sample vel commands from min [-1] or zero [0] or max [1] range only
         stop_heading_at_limit = True # stop heading updates when vel is limited
-        dynamic_resample_commands = False # sample commands with low bounds
+        dynamic_resample_commands = True # sample commands with low bounds
         command_range_curriculum = []
         # command_range_curriculum = [{ # list for command range curriculums at specific training iterations
         #     'iter': 20000, # training iteration at which the command ranges are updated

legged_gym/envs/go2/go2_config_vanilla2.py

@@ -0,0 +1,324 @@
+# -*- coding: utf-8 -*-
+'''
+@File    : go2_config_vanilla2.py
+@Time    : 2026/01/10 02:27:28
+@Author  : wty-yy
+@Version : 1.0
+@Blog    : https://wty-yy.github.io/
+@Desc    : Go2 vanilla2 training config, same as unitree rl gym except domain randomization and rewards
+episode length 20, resample commands 10 sec,
+open heading_command
+close move_down_by_accumulated_xy_command, dynamic_resample_commands, zero_command_curriculum, limit_vel_prob, command_range_curriculum, dynamic_sigma
+'''
+import math
+from legged_gym.envs.base.legged_robot_config import LeggedRobotCfg, LeggedRobotCfgPPO, LeggedRobotCfgCTS, LeggedRobotCfgMoECTS, LeggedRobotCfgMoECTS, LeggedRobotCfgMCPCTS, LeggedRobotCfgACMoECTS, LeggedRobotCfgDualMoECTS, LeggedRobotCfgREMCTS
+
+class GO2Cfg(LeggedRobotCfg):
+    class init_state(LeggedRobotCfg.init_state):
+        pos = [0.0, 0.0, 0.42] # x,y,z [m]
+        default_joint_angles = { # = target angles [rad] when action = 0.0
+            'FL_hip_joint': 0.1,   # [rad]
+            'RL_hip_joint': 0.1,   # [rad]
+            'FR_hip_joint': -0.1 ,  # [rad]
+            'RR_hip_joint': -0.1,   # [rad]
+
+            'FL_thigh_joint': 0.8,     # [rad]
+            'RL_thigh_joint': 1.,   # [rad]
+            'FR_thigh_joint': 0.8,     # [rad]
+            'RR_thigh_joint': 1.,   # [rad]
+
+            'FL_calf_joint': -1.5,   # [rad]
+            'RL_calf_joint': -1.5,    # [rad]
+            'FR_calf_joint': -1.5,  # [rad]
+            'RR_calf_joint': -1.5,    # [rad]
+        }
+        turn_over = False # initialize the robot in a flipped over position
+        # turn_over_proportions = [0.1, 0.3, 0.6] # proportions for backflip, sideflip, noflip
+        turn_over_proportions = [0.0, 0.2, 0.8] # proportions for backflip, sideflip, noflip
+        turn_over_init_heights = { # initial heights range for each flip type
+            'backflip': [0.10, 0.15],
+            'sideflip': [0.16, 0.21],
+        }
+        # turn_over_proportions = [0.0, 1.0, 0.0] # proportions for backflip, sideflip, noflip
+
+    class env(LeggedRobotCfg.env):
+        num_envs = 8192
+        num_observations = 45
+        # obs(45) + base_lin_vel(3) + height_measurements(187)
+        num_privileged_obs = 45 + 3 + 4 + 12 + 12 + 187  # 263
+        # num_privileged_obs = 45 + 3 + 187  # 235
+        # num_privileged_obs = 48  # without height measurements
+        episode_length_s = 20
+
+    class domain_rand(LeggedRobotCfg.domain_rand):
+        ### Robot properties ###
+        randomize_friction = True
+        friction_range = [0.0, 2.0]
+
+        randomize_base_mass = True
+        added_mass_range = [-1., 1.]
+
+        randomize_link_mass = True
+        multiplied_link_mass_range = [0.9, 1.1]
+
+        randomize_base_com = True
+        added_base_com_range = [-0.03, 0.03]
+
+        randomize_restitution = True # restitution to robot links (Robot init)
+        restitution_range = [0.0, 0.5]
+
+        ### Environment reset ###
+        randomize_pd_gains = True
+        stiffness_multiplier_range = [0.9, 1.1]  
+        damping_multiplier_range = [0.9, 1.1]    
+
+        randomize_motor_zero_offset = True
+        motor_zero_offset_range = [-0.035, 0.035]
+
+        randomize_motor_strength = True # (Env reset)
+        motor_strength_range = [0.8, 1.2]
+
+        ### Environment step ###
+        push_robots = True
+        push_interval_s = 4
+        max_push_vel_xy = 0.4
+        max_push_ang_vel = 0.6
+
+        randomize_action_delay = True # use last_action with 0~20 ms delay, 4 decimation
+
+    class control(LeggedRobotCfg.control):
+        # PD Drive parameters:
+        control_type = 'P'
+        stiffness = {'joint': 20.0}  # [N*m/rad]
+        damping = {'joint': 0.5}     # [N*m*s/rad]
+        # action scale: target angle = actionScale * action + defaultAngle
+        action_scale = 0.25
+        # decimation: Number of control action updates @ sim DT per policy DT
+        decimation = 4
+    
+    class terrain(LeggedRobotCfg.terrain):
+        max_init_terrain_level = 5
+        # [wave, slope, rough_slope, stairs up, stairs down, obstacles, stepping_stones, gap, flat]
+        # terrain_proportions = [0.2, 0.05, 0.05, 0.30, 0.05, 0.25, 0.0, 0.0, 0.1]  # 更偏向wave
+        terrain_proportions = [0.05, 0.20, 0.05, 0.25, 0.10, 0.20, 0.0, 0.0, 0.15]  # 这个更偏向平地斜坡
+        # terrain_proportions = [0.20, 0.05, 0.05, 0.30, 0.15, 0.20, 0.0, 0.0, 0.05]  # 更偏向wave和stairs
+        # terrain_proportions = [0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0]
+        # terrain_proportions = [0.3, 0.3, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1]
+        # terrain_proportions = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0]
+        move_down_by_accumulated_xy_command = False # move down the terrain curriculum based on accumulated xy command distance instead of absolute distance
+        
+    class commands(LeggedRobotCfg.commands):
+        curriculum = False
+        max_curriculum = 1.
+        num_commands = 4 # default: lin_vel_x, lin_vel_y, ang_vel_yaw (in heading mode ang_vel_yaw is recomputed from heading error)
+        resampling_time = 10. # time before command are changed[s]
+        heading_command = True # if true: compute ang vel command from heading error
+        # start training with zero commands and then gradually increase zero command probability
+        zero_command_curriculum = None
+        # zero_command_curriculum = {'start_iter': 0, 'end_iter': 1500, 'start_value': 0.0, 'end_value': 0.1}
+        limit_ang_vel_at_zero_command_prob = 0.0 # probability of add limiting angular velocity commands when zero command is sampled
+        limit_vel_prob = 0.0 # probability of limiting linear velocity command
+        limit_vel_invert_when_continuous = True # invert the limit logic when using continuous sample limit velocity commands
+        limit_vel = {"lin_vel_x": [-1, 1], "lin_vel_y": [-1, 1], "ang_vel_yaw": [-1, 0, 1]} # sample vel commands from min [-1] or zero [0] or max [1] range only
+        stop_heading_at_limit = True # stop heading updates when vel is limited
+        dynamic_resample_commands = False # sample commands with low bounds
+        command_range_curriculum = []
+        # command_range_curriculum = [{ # list for command range curriculums at specific training iterations
+        #     'iter': 20000, # training iteration at which the command ranges are updated
+        #     'lin_vel_x': [-1.0, 1.0], # min max [m/s]
+        #     'lin_vel_y': [-1.0, 1.0], # min max [m/s]
+        #     'ang_vel_yaw': [-1.5, 1.5], # min max [rad/s]
+        #     'heading': [-1.57, 1.57], # min max [rad]
+        # }, { # list for command range curriculums at specific training iterations
+        #     'iter': 50000, # training iteration at which the command ranges are updated
+        #     'lin_vel_x': [-2.0, 2.0], # min max [m/s]
+        #     'lin_vel_y': [-1.0, 1.0], # min max [m/s]
+        #     'ang_vel_yaw': [-2.0, 2.0], # min max [rad/s]
+        #     'heading': [-1.57, 1.57], # min max [rad]
+        # }]
+        turn_over_zero_time = { # if turn_over is true, time robot must be stable before sampling new commands after a turn over
+            "backflip": 5.0,
+            "sideflip": 3.0,
+        }
+        # [wave, slope, rough slope, stairs up, stairs down, obstacles, stepping stones, gap, flat]
+        terrain_max_command_ranges = [
+            {'lin_vel_x': [-1.5, 1.5], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # wave
+            {'lin_vel_x': [-1.5, 1.5], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # slope
+            {'lin_vel_x': [-1.5, 1.5], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # rough slope
+            {'lin_vel_x': [-1.0, 1.0], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # stairs up
+            {'lin_vel_x': [-1.0, 1.0], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # stairs down
+            {'lin_vel_x': [-1.0, 1.0], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # obstacles
+            {'lin_vel_x': [-1.0, 1.0], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # stepping stones
+            {'lin_vel_x': [-1.0, 1.0], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-1.5, 1.5], 'heading': [-1.57, 1.57]},  # gap
+            {'lin_vel_x': [-2.0, 2.0], 'lin_vel_y': [-1.0, 1.0], 'ang_vel_yaw': [-2.0, 2.0], 'heading': [-1.57, 1.57]},  # flat
+        ]
+
+        class ranges:
+            lin_vel_x = [-2.0, 2.0] # min max [m/s]
+            lin_vel_y = [-1.0, 1.0] # min max [m/s]
+            ang_vel_yaw = [-2.0, 2.0]   # min max [rad/s]
+            heading = [-1.57, 1.57] # min max [rad]
+        
+    class asset(LeggedRobotCfg.asset):
+        file = '{LEGGED_GYM_ROOT_DIR}/resources/robots/go2/urdf/go2.urdf'
+        name = "go2"
+        foot_name = "foot"
+        penalize_contacts_on = ["thigh", "calf"]
+        terminate_after_contacts_on = ["base"]
+        self_collisions = 1 # 1 to disable, 0 to enable...bitwise filter
+  
+    class rewards(LeggedRobotCfg.rewards):
+        soft_dof_pos_limit = 0.9
+        base_height_target = 0.38
+        only_positive_rewards = False
+        max_contact_force = 147. # forces above this value are penalized, go2 weight 15kg
+        curriculum_rewards = [
+            {'reward_name': 'lin_vel_z', 'start_iter': 0, 'end_iter': 1500, 'start_value': 1.0, 'end_value': 0.0},
+            {'reward_name': 'correct_base_height', 'start_iter': 0, 'end_iter': 5000, 'start_value': 1.0, 'end_value': 10.0},
+            # {'reward_name': 'dof_power', 'start_iter': 0, 'end_iter': 3000, 'start_value': 1.0, 'end_value': 0.1},
+            # {'reward_name': 'upright', 'start_iter': 0, 'end_iter': 1500, 'start_value': 1.0, 'end_value': 0.0},
+        ]
+        tracking_sigma = 0.25 # tracking reward = exp(-error^2/sigma)
+        dynamic_sigma = None
+        # dynamic_sigma = { # linear interpolation of sigma based on command velocity, **Must start terrain curriculum first**
+        #     "min_lin_vel": 0.5, # min abs linear velocity to have default sigma
+        #     "max_lin_vel": 1.5, # max abs linear velocity to have max sigma
+        #     "min_ang_vel": 1.0, # min abs angular velocity to have default sigma
+        #     "max_ang_vel": 2.0, # max abs angular velocity to have max sigma
+        #     # wave, slope, rough_slope, stairs up, stairs down, obstacles, stepping_stones, gap, flat]
+        #     # "max_sigma": [1/3, 1/4, 1/4, 1/2.7, 1/2.7, 1/2, 1, 1, 1/4]
+        #     "max_sigma": [5/12, 1/4, 1/4, 1/2, 1/2, 3/4, 1, 1, 1/4]
+        # }
+        min_legs_distance = 0.1  # min distance between legs to not be considered stumbling
+        class scales:
+            # tracking_lin_vel = 1.0
+            # tracking_ang_vel = 0.2
+            # lin_vel_z = -10.0
+            # base_height = -50.0
+            # action_rate = -0.005
+            # similar_to_default = -0.1
+            # dof_power = -1e-3  # 能够明显抑制跳跃
+            # dof_acc = -3e-7
+
+            # tracking_lin_vel = 1.0
+            # tracking_ang_vel = 0.5
+            # lin_vel_z = -2.0
+            # ang_vel_xy = -0.05
+            # dof_acc = -2.5e-7
+            # dof_power = -1e-3  # 能够明显抑制跳跃
+            # # torques = -1e-4  # 无用会走着走着倒了
+            # correct_base_height = -10.0
+            # action_rate = -0.01
+            # action_smoothness = -0.01
+            # collision = -1.0
+            # dof_pos_limits = -2.0
+            # feet_regulation = -0.05
+            # hip_to_default = -0.1
+            # similar_to_default = -0.05
+
+            # CTS reward
+            tracking_lin_vel = 1.0
+            tracking_ang_vel = 0.5
+            lin_vel_z = -2.0
+            ang_vel_xy = -0.05
+            dof_acc = -2.5e-7
+            dof_power = -2e-5
+            torques = -1e-4
+            correct_base_height = -1.0
+            action_rate = -0.01
+            action_smoothness = -0.01
+            collision = -1.0
+            dof_pos_limits = -2.0
+            feet_regulation = -0.05
+            # CTS奖励训出来双脚距离非常近, 真机效果很差, 但是sim2sim能上20cm楼梯, 尝试加入hip_to_default奖励或similar_to_default奖励
+            hip_to_default = -0.05  # 在训练到y=1.5时, 双脚会明显碰撞, 为避免该问题提升hip, 效果更差, 还是保持0.05 (y最大也只到0.1了)
+            # legs_distance = -1.5  # 奖励双脚距离, 避免CTS训练出来双脚距离过近, 尝试加入后robogauge flat验证效果变差, 删除
+            # similar_to_default = -0.01
+            # feet_contact_forces = -1.0  # 尝试加入但并没有起到任何效果, 删除
+
+        turn_over_roll_threshold = math.pi / 4 # threshold on roll to use turn over rewards
+        class turn_over_scales:
+            upright = 1.0
+            # dof_acc = -2.5e-7
+            # dof_power = -2e-5
+            # action_rate = -0.001
+            # action_smoothness = -0.001
+
+    class noise(LeggedRobotCfg.noise):
+        add_noise = True
+
+class GO2CfgPPO(LeggedRobotCfgPPO):
+    class algorithm(LeggedRobotCfgPPO.algorithm):
+        entropy_coef = 0.01
+    class runner(LeggedRobotCfgPPO.runner):
+        run_name = ''
+        experiment_name = 'go2_ppo'
+        max_iterations = 100000
+        save_interval = 500
+
+class GO2CfgCTS(LeggedRobotCfgCTS):
+    class runner(LeggedRobotCfgCTS.runner):
+        num_steps_per_env = 24
+        run_name = ''
+        experiment_name = 'go2_cts'
+        max_iterations = 150000
+        save_interval = 500
+    
+    class policy(LeggedRobotCfgCTS.policy):
+        latent_dim = 32
+        norm_type = 'l2norm'
+
+class GO2CfgMoECTS(LeggedRobotCfgMoECTS):
+    class policy(LeggedRobotCfgMoECTS.policy):
+        obs_no_goal_mask = [True] * 6 + [False] * 3 + [True] * 36  # mask for obs without command info
+        student_expert_num = 8 # number of experts in the student model
+    
+    class algorithm(LeggedRobotCfgMoECTS.algorithm):
+        load_balance_coef = 0.01
+
+    class runner(LeggedRobotCfgMoECTS.runner):
+        run_name = ''
+        experiment_name = 'go2_moe_cts'
+        max_iterations = 150000
+        save_interval = 500
+
+class GO2CfgMCPCTS(LeggedRobotCfgMCPCTS):
+    class policy(LeggedRobotCfgMCPCTS.policy):
+        obs_no_goal_mask = [True] * 6 + [False] * 3 + [True] * 36  # mask for obs without command info
+        student_expert_num = 8 # number of experts in the student model
+
+    class runner(LeggedRobotCfgMCPCTS.runner):
+        run_name = ''
+        experiment_name = 'go2_mcp_cts'
+        max_iterations = 150000
+        save_interval = 500
+
+class GO2CfgACMoECTS(LeggedRobotCfgACMoECTS):
+    class policy(LeggedRobotCfgACMoECTS.policy):
+        expert_num = 8  # number of experts in the student model
+    
+    class runner(LeggedRobotCfgACMoECTS.runner):
+        run_name = ''
+        experiment_name = 'go2_ac_moe_cts'
+        max_iterations = 150000
+        save_interval = 500
+
+class GO2CfgDualMoECTS(LeggedRobotCfgDualMoECTS):
+    class policy(LeggedRobotCfgDualMoECTS.policy):
+        expert_num = 8  # number of experts in the student model
+    
+    class runner(LeggedRobotCfgDualMoECTS.runner):
+        run_name = ''
+        experiment_name = 'go2_dual_moe_cts'
+        max_iterations = 150000
+        save_interval = 500
+
+class GO2CfgREMCTS(LeggedRobotCfgREMCTS):
+    class policy(LeggedRobotCfgREMCTS.policy):
+        expert_num = 8  # number of experts in the student model
+    
+    class runner(LeggedRobotCfgREMCTS.runner):
+        run_name = ''
+        experiment_name = 'go2_rem_cts'
+        max_iterations = 150000
+        save_interval = 500

legged_gym/utils/exporter.py

@@ -210,10 +210,13 @@ class _OnnxPolicyExporter(torch.nn.Module):
             
         elif hasattr(policy, "student_moe_encoder"):
             self.student_moe_encoder = copy.deepcopy(policy.student_moe_encoder)
-            self.obs_no_goal_mask = copy.deepcopy(policy.obs_no_goal_mask).cpu()
             self.history_length = policy.history.shape[1]
             self.forward = self.forward_moe_cts
             self.input_dim = self.history_length * policy.history.shape[2]
+            if hasattr(policy, "obs_no_goal_mask"):
+                self.obs_no_goal_mask = copy.deepcopy(policy.obs_no_goal_mask).cpu()
+            else:
+                self.forward = self.forward_rem_cts
         
         else:  # PPO
             self.forward = self.forward_ppo
@@ -287,6 +290,17 @@ class _OnnxPolicyExporter(torch.nn.Module):
 
         return self.actor(x), weights, latent
 
+    def forward_rem_cts(self, x):
+        x = self.normalizer(x)
+        history, obs_dim = self.flatten_obs(x)
+
+        last_obs = history[:, -obs_dim:]
+
+        latent, weights = self.student_moe_encoder(history)
+        x = torch.cat([latent, last_obs], dim=1)
+
+        return self.actor(x), weights, latent
+    
     def forward_mcp_cts(self, x):
         x = self.normalizer(x)
         history, obs_dim = self.flatten_obs(x)

setup.py

@@ -2,7 +2,7 @@ from setuptools import find_packages
 from distutils.core import setup
 
 setup(name='go2_rl_gym',
-      version='0.1.7',
+      version='0.1.8',
       author='Wu Tianyang',
       license="MIT",
       packages=find_packages(),