#!/usr/bin/env python3
"""MuJoCo sim2sim for go1-stairs-terrain-walk-no-linevel (57-dim obs, no linvel).

Loads the ONNX policy exported by export_go1_no_linevel_onnx.py and runs
inference in MuJoCo with PD control.

Usage:
    # Default: combined flat+rough+stairs terrain, random spawn
    uv run scripts/go1_no_linevel_sim2sim_mujoco.py

    # Specific terrain
    uv run scripts/go1_no_linevel_sim2sim_mujoco.py --terrain flat
    uv run scripts/go1_no_linevel_sim2sim_mujoco.py --terrain rough
    uv run scripts/go1_no_linevel_sim2sim_mujoco.py --terrain stairs

    # Custom ONNX path
    uv run scripts/go1_no_linevel_sim2sim_mujoco.py --onnx ./exports_go1_no_linevel/policy.onnx

Keyboard controls:
    W/S     - forward/backward
    A/D     - turn left/right
    Q/E     - strafe left/right
    Space   - stop
    R       - reset robot
    1/2/3   - switch terrain (flat/rough/stairs)
    Esc     - quit
"""

import numpy as np
import mujoco
from mujoco import viewer
import os
import threading
import signal
import queue
import argparse
import time

g_exit_requested = False


def signal_handler(signum, frame):
    global g_exit_requested
    g_exit_requested = True


signal.signal(signal.SIGINT, signal_handler)

# ============================================================
# Paths
# ============================================================
HERE = os.path.dirname(os.path.abspath(__file__))
DEFAULT_ONNX_PATH = os.path.join(HERE, "better3.onnx")
GO1_XML = os.path.join(HERE, "..", "sim2sim_mujoco_example", "data", "go1", "xml", "go1.xml")

# ============================================================
# MotrixLab parameters (matching cfg.py + walk_stairs_terrain_no_linevel.py)
# ============================================================
NUM_OBS = 57  # 57-dim: NO linear velocity, WITH contact forces
NUM_ACTIONS = 12
OBS_SCALES = {"ang_vel": 0.25, "dof_pos": 1.0, "dof_vel": 0.05}
ACTION_SCALE = 0.05
KP, KD = 80.0, 1.0
CLIP_ACTIONS = 23.7
CLIP_OBSERVATIONS = 100.0
MAX_LIN_VEL_X = 1.0
MAX_LIN_VEL_Y = 1.0
MAX_ANG_VEL = 1.0

# ============================================================
# Joint names and order
# ============================================================
POLICY_JOINT_NAMES = [
    "FR_hip", "FR_thigh", "FR_calf",
    "FL_hip", "FL_thigh", "FL_calf",
    "RR_hip", "RR_thigh", "RR_calf",
    "RL_hip", "RL_thigh", "RL_calf",
]

DEFAULT_JOINT_ANGLES = np.array([
    -0.0, 0.9, -1.8,
    0.0, 0.9, -1.8,
    -0.0, 0.9, -1.8,
    0.0, 0.9, -1.8,
], dtype=np.float32)

FEET = ["FR", "FL", "RR", "RL"]

# Terrain spawn positions (world Y)
TERRAIN_SPAWN = {
    "flat": np.array([0.0, 54.0, 0.42], dtype=np.float64),
    "rough": np.array([0.0, 32.0, 0.42], dtype=np.float64),
    "stairs": np.array([0.0, 0.0, 0.42], dtype=np.float64),
}

# ============================================================
# Keyboard input
# ============================================================
from pynput import keyboard


class KeyboardReader:
    def __init__(self):
        self._event_queue = queue.Queue()
        self.running = True
        self.shared_keys_held = set()
        self.shared_one_shot = set()
        self._reader_thread = None
        self._listener = None

    def _normalize_key(self, key):
        try:
            if hasattr(key, "char") and key.char is not None:
                return key.char.lower()
        except Exception:
            pass
        key_str = str(key)
        if key_str == "Key.esc":
            return "escape"
        elif key_str == "Key.space":
            return "space"
        elif key_str.startswith("Key."):
            return key_str.lower()
        return key_str.lower()

    def _reader_worker(self):
        while self.running:
            try:
                event_type, key = self._event_queue.get(timeout=0.05)
                k = self._normalize_key(key)
                if event_type == "press":
                    self.shared_keys_held.add(k)
                    self.shared_one_shot.discard(k)
                elif event_type == "release":
                    self.shared_keys_held.discard(k)
                    self.shared_one_shot.discard(k)
            except queue.Empty:
                pass

    def init(self):
        def on_press(key):
            self._event_queue.put(("press", key))

        def on_release(key):
            self._event_queue.put(("release", key))

        try:
            self._listener = keyboard.Listener(on_press=on_press, on_release=on_release)
            self._listener.start()
            self._reader_thread = threading.Thread(target=self._reader_worker, daemon=True)
            self._reader_thread.start()
            print("[INFO] Keyboard listener started")
        except Exception as e:
            print(f"[WARN] Cannot init keyboard: {e}")

    def is_key_pressed(self, key):
        k = self._normalize_key(key) if isinstance(key, str) else self._normalize_key(key)
        if k not in self.shared_keys_held or k in self.shared_one_shot:
            return False
        self.shared_one_shot.add(k)
        return True

    def is_key_held(self, key):
        k = self._normalize_key(key) if isinstance(key, str) else self._normalize_key(key)
        return k in self.shared_keys_held

    def restore(self):
        self.running = False
        if self._listener:
            self._listener.stop()


# ============================================================
# Sensor reading
# ============================================================
def get_sensor(model, data, name):
    sid = mujoco.mj_name2id(model, mujoco.mjtObj.mjOBJ_SENSOR, name)
    if sid < 0:
        return None
    adr = model.sensor_adr[sid]
    dim = model.sensor_dim[sid]
    return data.sensordata[adr : adr + dim].copy()


def read_contact_forces(model, data, base_rot):
    """Read foot contact forces (12-dim, body frame) from MuJoCo contact sensors.

    Tries _stairs, _rough, _flat suffixes for each foot, picking the first
    sensor that returns non-zero data. In MuJoCo, `data="force"` returns a
    scalar (normal force). We construct a 3D force vector by projecting onto
    the body-frame Z axis as an approximation.
    """
    forces = np.zeros(12, dtype=np.float32)
    for i, foot in enumerate(FEET):
        f_scalar = 0.0
        for suffix in ["_stairs", "_rough", "_flat"]:
            name = f"{foot}_foot_contact{suffix}"
            v = get_sensor(model, data, name)
            if v is not None and np.abs(v[0]) > 1e-6:
                f_scalar = v[0]
                break
        # Assume contact force is approximately vertical (world Z),
        # rotate into body frame
        force_world = np.array([0.0, 0.0, f_scalar], dtype=np.float64)
        force_body = base_rot.T @ force_world
        forces[i * 3 : i * 3 + 3] = force_body.astype(np.float32)
    return forces


def compute_observations(model, data, commands, last_actions, base_rot):
    """Compute 57-dim observation matching go1-stairs-terrain-walk-no-linevel.

    Layout (57 dims, NO linear velocity):
      [0:3]   gyro (ang_vel * 0.25)
      [3:6]   gravity vector (body frame)
      [6:18]  joint angle deviation (dof_pos * 1.0)
      [18:30] joint velocity (dof_vel * 0.05)
      [30:42] last actions (raw)
      [42:45] commands [vx*2.0, vy*2.0, wz*0.25]
      [45:57] foot contact forces (body frame, raw)
    """
    obs = np.zeros(NUM_OBS, dtype=np.float32)

    # Gyro
    gyro = get_sensor(model, data, "gyro")
    if gyro is not None:
        obs[0:3] = gyro * OBS_SCALES["ang_vel"]
    else:
        obs[0:3] = data.qvel[3:6] * OBS_SCALES["ang_vel"]

    # Gravity vector (body frame)
    gravity_world = np.array([0.0, 0.0, -1.0], dtype=np.float64)
    local_gravity = base_rot.T @ gravity_world
    obs[3:6] = local_gravity.astype(np.float32)

    # Joint position deviation
    joint_pos = data.qpos[7:19]
    dof_pos_rel = (joint_pos - DEFAULT_JOINT_ANGLES) * OBS_SCALES["dof_pos"]
    obs[6:18] = dof_pos_rel

    # Joint velocity
    joint_vel = data.qvel[6:18]
    obs[18:30] = joint_vel * OBS_SCALES["dof_vel"]

    # Last actions
    obs[30:42] = last_actions

    # Commands (scale matching MotrixLab: [2.0, 2.0, 0.25])
    obs[42] = commands[0] * 2.0
    obs[43] = commands[1] * 2.0
    obs[44] = commands[2] * 0.25

    # Contact forces
    # obs[45:57] = read_contact_forces(model, data, base_rot)  # disabled: test with zeros
    obs[45:57] = np.zeros(12, dtype=np.float32)

    obs = np.clip(obs, -CLIP_OBSERVATIONS, CLIP_OBSERVATIONS)
    return obs


# ============================================================
# Main
# ============================================================
def main():
    import onnxruntime as ort

    parser = argparse.ArgumentParser(description="MotrixLab Go1 No-Linevel Policy Inference in MuJoCo")
    parser.add_argument("--onnx", type=str, default=DEFAULT_ONNX_PATH)
    parser.add_argument(
        "--terrain", type=str, default="combined",
        choices=["flat", "rough", "stairs", "combined"],
        help="Terrain type (combined = flat+rough+stairs in one scene)",
    )
    args = parser.parse_args()

    # Use local Go1 XML (terrain switching disabled on macOS)
    os.chdir(os.path.dirname(GO1_XML))  # mesh paths are relative
    with open(GO1_XML, "r") as f:
        xml_content = f.read()

    model = mujoco.MjModel.from_xml_string(xml_content)
    data = mujoco.MjData(model)

    print(f"[INFO] Terrain: {args.terrain}")
    print(f"[INFO] Model: {model.nbody} bodies, {model.nq} DoF, {model.nu} actuators")
    print(f"[INFO] Timestep: {model.opt.timestep}")

    # Use simple spawn
    spawn_xyz = np.array([0.0, 0.0, 0.42], dtype=np.float64)

    data.qpos[0:3] = spawn_xyz
    data.qpos[3:7] = np.array([1.0, 0.0, 0.0, 0.0])
    data.qpos[7:19] = DEFAULT_JOINT_ANGLES
    data.qvel[:] = 0.0
    data.ctrl[:] = 0.0
    mujoco.mj_forward(model, data)

    # Load ONNX
    session = ort.InferenceSession(args.onnx, providers=["CPUExecutionProvider"])
    print(f"[INFO] ONNX loaded: {args.onnx}")

    # Main loop
    ctrl_dt = 0.01
    num_steps_per_inference = int(ctrl_dt / model.opt.timestep)
    print(f"[INFO] Inference every {num_steps_per_inference} sim steps")

    step_count = 0
    inference_step = 0
    commands = np.zeros(3, dtype=np.float32)
    last_actions = np.zeros(NUM_ACTIONS, dtype=np.float32)
    action = np.zeros(NUM_ACTIONS, dtype=np.float32)

    keyboard_reader = KeyboardReader()
    keyboard_reader.init()

    viewer_handle = viewer.launch_passive(model, data)
    print("[INFO] Viewer launched!")
    print("[KEYS] WASD=move, QE=strafe, Space=stop, R=reset, 1/2/3=terrain, Esc=quit")

    loop_start_time = time.time()

    while viewer_handle.is_running() and not g_exit_requested:
        # --- Keyboard input ---
        x_vel, y_vel, yaw_vel = 0.0, 0.0, 0.0

        if keyboard_reader.is_key_held("w"):
            x_vel = MAX_LIN_VEL_X
        elif keyboard_reader.is_key_held("s"):
            x_vel = -MAX_LIN_VEL_X

        if keyboard_reader.is_key_held("q"):
            y_vel = MAX_LIN_VEL_Y
        elif keyboard_reader.is_key_held("e"):
            y_vel = -MAX_LIN_VEL_Y

        if keyboard_reader.is_key_held("a"):
            yaw_vel = MAX_ANG_VEL
        elif keyboard_reader.is_key_held("d"):
            yaw_vel = -MAX_ANG_VEL

        if keyboard_reader.is_key_pressed("space"):
            x_vel = y_vel = yaw_vel = 0.0

        # Reset
        if keyboard_reader.is_key_pressed("r"):
            spawn_xyz[:] = [0.0, 0.0, 0.42]

            data.qpos[0:3] = spawn_xyz
            data.qpos[3:7] = np.array([1.0, 0.0, 0.0, 0.0])
            data.qpos[7:19] = DEFAULT_JOINT_ANGLES
            data.qvel[:] = 0.0
            data.ctrl[:] = 0.0
            last_actions = np.zeros(NUM_ACTIONS, dtype=np.float32)
            mujoco.mj_forward(model, data)
            print(f"[RESET] pos={spawn_xyz}")

        if keyboard_reader.is_key_pressed("escape"):
            break

        # --- Inference ---
        if inference_step == 0:
            commands[0] = x_vel
            commands[1] = y_vel
            commands[2] = yaw_vel

            base_rot = data.xmat[1].reshape(3, 3)
            obs = compute_observations(model, data, commands, last_actions, base_rot)

            action = session.run(None, {"observations": obs.reshape(1, -1).astype(np.float32)})[0][0]
            action = np.clip(action, -CLIP_ACTIONS, CLIP_ACTIONS)
            last_actions = action.copy()

        # --- PD control ---
        joint_targets = DEFAULT_JOINT_ANGLES + action * ACTION_SCALE
        current_pos = data.qpos[7:19]
        current_vel = data.qvel[6:18]
        torques = KP * (joint_targets - current_pos) - KD * current_vel
        torques = np.clip(torques, -CLIP_ACTIONS, CLIP_ACTIONS)
        data.ctrl[:] = torques

        mujoco.mj_step(model, data)
        viewer_handle.sync()

        expected_time = step_count * ctrl_dt
        elapsed = time.time() - loop_start_time
        sleep_time = expected_time - elapsed
        if sleep_time > 0:
            time.sleep(sleep_time)

        step_count += 1
        inference_step = (inference_step + 1) % num_steps_per_inference

        if step_count % 500 == 0:
            trunk_z = data.qpos[2]
            print(f"[{step_count}] cmd=({x_vel:.1f},{y_vel:.1f},{yaw_vel:.1f}) "
                  f"z={trunk_z:.3f}m")

    keyboard_reader.restore()
    viewer_handle.close()


if __name__ == "__main__":
    main()