#!/usr/bin/env python3
"""DreamWaQ MuJoCo sim2sim — VAE encoder + Actor, 5-frame history buffer.

Usage:
    uv run scripts/dreamwaq_sim2sim_mujoco.py                          # flat
    uv run scripts/dreamwaq_sim2sim_mujoco.py --terrain rough
    uv run scripts/dreamwaq_sim2sim_mujoco.py --onnx path/to/policy.onnx

Controls:
    W/S: forward/back    Q/E: left/right    A/D: rotate
    Space: stop    R: reset    Esc: quit
"""
import numpy as np
import mujoco
from mujoco import viewer
import onnxruntime as ort
import os, sys, threading, queue, argparse, time, signal

g_exit_requested = False
signal.signal(signal.SIGINT, lambda *a: globals().update(g_exit_requested=True))

# ═══════════════════════════════════════════════════════════════════════
_PROJECT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
XML_DIR = os.path.join(_PROJECT, "motrix_envs", "src", "motrix_envs", "locomotion", "go1", "xmls")
DEFAULT_ONNX = os.path.join(_PROJECT, "exports_go1_dreamwaq", "policy.onnx")

# ── DreamWaQ params (matching training: PD 28/0.7, action_scale 0.25, ctrl_dt=0.02) ──
NUM_OBS = 45
NUM_ACTIONS = 12
HISTORY_LEN = 5
ACTION_SCALE = 0.25
KP = 28.0
KD = 0.7
CLIP_ACTIONS = 4.0
CLIP_TORQUES = 80.0
CLIP_OBS = 100.0
MAX_VX, MAX_VY, MAX_WZ = 1.0, 1.0, 1.0

# DreamWaQ default joint angles — MUST match MuJoCo XML joint order:
# qpos[7:19] = FR_hip,FR_thigh,FR_calf, FL_hip,FL_thigh,FL_calf, RR_hip,RR_thigh,RR_calf, RL_hip,RL_thigh,RL_calf
# 必须与 MotrixSim 训练的 default_angles 完全一致！
DEFAULT_ANGLES = np.array([
    0.0, 0.9, -1.8,   # FR
    0.0, 0.9, -1.8,   # FL
    0.0, 0.9, -1.8,   # RR
    0.0, 0.9, -1.8,   # RL
], dtype=np.float32)


# ═══════════════════════════════════════════════════════════════════════
# Keyboard
# ═══════════════════════════════════════════════════════════════════════
from pynput import keyboard

class KB:
    def __init__(self):
        self._q = queue.Queue(); self.running = True
        self.held = set(); self._t = None; self._l = None
    def _n(self, k):
        try:
            if hasattr(k, 'char') and k.char: return k.char.lower()
        except: pass
        return str(k).lower()
    def _w(self):
        while self.running:
            try:
                et, k = self._q.get(timeout=0.05)
                n = self._n(k)
                if et == 'press': self.held.add(n)
                elif et == 'release': self.held.discard(n)
            except queue.Empty: pass
    def init(self):
        self._l = keyboard.Listener(on_press=lambda k: self._q.put(('press', k)),
                                     on_release=lambda k: self._q.put(('release', k)))
        self._l.start()
        self._t = threading.Thread(target=self._w, daemon=True); self._t.start()
    def stop(self): self.running = False; self._l.stop()


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


def compute_obs(model, data, commands, last_action):
    """DreamWaQ observation (Manaro-Alpha order):
    ang_vel(3) + gravity(3) + commands(3) + joint_pos(12) + joint_vel(12) + actions(12) = 45
    """
    obs = np.zeros(NUM_OBS, dtype=np.float32)
    # ang_vel [0:3]
    g = get_sensor(model, data, "gyro")
    obs[0:3] = (g if g is not None else data.qvel[3:6]) * 0.25
    # gravity [3:6] (read from MuJoCo model, matching training)
    grav_world = model.opt.gravity.copy()
    grav_world = grav_world / np.linalg.norm(grav_world)  # normalize
    R = data.xmat[1].reshape(3, 3)
    obs[3:6] = (R.T @ grav_world).astype(np.float32)
    # commands [6:9]
    obs[6:9] = commands * np.array([2.0, 2.0, 0.25], dtype=np.float32)
    # joint_pos [9:21]
    obs[9:21] = (data.qpos[7:19] - DEFAULT_ANGLES) * 1.0
    # joint_vel [21:33]
    obs[21:33] = data.qvel[6:18] * 0.05
    # last_action [33:45]
    obs[33:45] = last_action
    return np.clip(obs, -CLIP_OBS, CLIP_OBS)


# ═══════════════════════════════════════════════════════════════════════
# Main
# ═══════════════════════════════════════════════════════════════════════
def main():
    p = argparse.ArgumentParser()
    p.add_argument("--onnx", default=DEFAULT_ONNX)
    p.add_argument("--terrain", default="flat", choices=["flat", "rough", "stairs", "dreamwaq", "stairs_test", "stairs_box", "flat_stairs"])
    p.add_argument("--level", type=int, default=0, help="terrain difficulty level 0-9 (0=flat, 9=hardest)")
    args = p.parse_args()

    # Select XML scene
    terrain_map = {
        "flat": "scene_dreamwaq_flat.xml",
        "rough": "scene_rough_terrain.xml",
        "stairs": "scene_stairs_terrain.xml",
        "dreamwaq": "scene_dreamwaq_terrain.xml",
        "stairs_test": "scene_stairs_test.xml",
        "stairs_box": "scene_stairs_box.xml",
        "flat_stairs": "scene_flat_stairs.xml",
    }
    xml_file = os.path.join(XML_DIR, terrain_map[args.terrain])

    if not os.path.exists(args.onnx):
        print(f"[ERROR] ONNX not found: {args.onnx}")
        print("Run: uv run scripts/export_dreamwaq_onnx.py (after training completes)")
        sys.exit(1)

    os.chdir(XML_DIR)
    with open(xml_file) as f:
        model = mujoco.MjModel.from_xml_string(f.read())
    data = mujoco.MjData(model)

    # Spawn pose. Hfield heights: MuJoCo z = gp[2] + sbase + (hd * ztop).
    # The stairs_test terrain has sbase=0, flat platform z=0; just lift by clearance.
    if args.terrain == "flat_stairs":
        lvl = max(0, min(1, args.level))
        col = np.random.randint(0, 4)
        spawn_y = 4.0 - lvl * 8.0     # level 0 flat at y=+4, level 1 stairs at y=-4
        spawn_x = -12.0 + col * 8.0   # platform center (cell center x)
    elif args.terrain == "stairs_test":
        spawn_x, spawn_y = -7.5, -4.0  # flat approach before first step (1m zone)
    elif args.terrain == "stairs_box":
        spawn_x, spawn_y = -2.0, 0.0   # flat ground before stairs
    elif args.terrain == "dreamwaq":
        lvl = max(0, min(9, args.level))
        col = np.random.randint(0, 4)   # NUM_COLS=4
        spawn_y = 36.0 - lvl * 8.0    # level 0 flat at y=+36, level 9 stairs at y=-36
        spawn_x = -12.0 + col * 8.0 + 4.0  # centre of cell
        print(f"[Level {lvl}] type={col}  spawn=({spawn_x:.1f}, {spawn_y:.1f})")
    else:
        spawn_x, spawn_y = (0.0, 0.0)

    # When DISPLAY is a virtual framebuffer (Xvfb), MuJoCo headless rendering is
    # handled transparently; on a real display this opens a normal GUI window.
    # No explicit headless flag needed — MuJoCo glfw detects the display type.

    def hfield_z(mx, my):
        gid = mujoco.mj_name2id(model, mujoco.mjtObj.mjOBJ_GEOM, "floor")
        if gid < 0 or model.geom_type[gid] != mujoco.mjtGeom.mjGEOM_HFIELD:
            return 0.0
        hf = model.geom_dataid[gid]
        nrow, ncol = int(model.hfield_nrow[hf]), int(model.hfield_ncol[hf])
        sx, sy, ztop, sbase = model.hfield_size[hf]
        adr = model.hfield_adr[hf]
        hd = model.hfield_data[adr:adr + nrow * ncol].reshape(nrow, ncol)
        gp = model.geom_pos[gid]
        col = int(np.clip(((mx - gp[0]) / sx * 0.5 + 0.5) * (ncol - 1), 0, ncol - 1))
        row = int(np.clip(((my - gp[1]) / sy * 0.5 + 0.5) * (nrow - 1), 0, nrow - 1))
        return float(gp[2] + sbase + hd[row, col] * ztop)

    spawn_z = hfield_z(spawn_x, spawn_y) + 0.45  # standing clearance above terrain

    def reset_state():
        data.qpos[:] = 0
        data.qpos[0:3] = [spawn_x, spawn_y, spawn_z]; data.qpos[3:7] = [1, 0, 0, 0]
        data.qpos[7:19] = DEFAULT_ANGLES; data.qvel[:] = 0
        mujoco.mj_forward(model, data)

    reset_state()

    # ONNX (2 inputs: observations + obs_history)
    session = ort.InferenceSession(args.onnx, providers=['CPUExecutionProvider'])
    print(f"[DreamWaQ] {args.onnx}")
    print(f"[Terrain] {args.terrain}")
    print(f"[CTRL] W/S前后 Q/E左右 A/D旋转 Space停 R重置 Esc退出")

    kb = KB(); kb.init()
    view = viewer.launch_passive(model, data)

    # Camera tracking: follow the trunk body
    trunk_id = mujoco.mj_name2id(model, mujoco.mjtObj.mjOBJ_BODY, "trunk")
    view.cam.lookat = data.body(trunk_id).xpos.copy()
    view.cam.type = mujoco.mjtCamera.mjCAMERA_TRACKING
    view.cam.trackbodyid = trunk_id

    step = 0
    vx, vy, wz = 0.0, 0.0, 0.0
    last_action = np.zeros(NUM_ACTIONS, dtype=np.float32)
    action = np.zeros(NUM_ACTIONS, dtype=np.float32)
    history = np.zeros((1, HISTORY_LEN, NUM_OBS), dtype=np.float32)
    decimation = 4  # MuJoCo dt=0.005, policy dt=0.02 (DreamWaQ aligned)

    loop_t0 = time.time()

    while view.is_running() and not g_exit_requested:
        keys = kb.held
        if 'escape' in keys: break
        if 'r' in keys:
            reset_state()
            last_action[:] = 0; history[:] = 0
            print("[R] Reset")
        if ' ' in keys: vx = vy = wz = 0.0

        vx = MAX_VX if 'w' in keys else (-MAX_VX if 's' in keys else 0.0)
        vy = MAX_VY if 'q' in keys else (-MAX_VY if 'e' in keys else 0.0)
        wz = MAX_WZ if 'a' in keys else (-MAX_WZ if 'd' in keys else 0.0)

        if step % decimation == 0:
            cmd = np.array([vx, vy, wz], dtype=np.float32)
            obs = compute_obs(model, data, cmd, last_action)

            # Shift history + add new obs
            history = np.concatenate([history[:, 1:, :], obs.reshape(1, 1, -1)], axis=1)

            # ONNX inference
            outputs = session.run(None, {
                'obs': obs.reshape(1, -1).astype(np.float32),
                'obs_history': history.reshape(1, -1).astype(np.float32),
            })
            action = outputs[0][0]
            action = np.clip(action, -CLIP_ACTIONS, CLIP_ACTIONS)
            last_action = action.copy()

        # PD control（对齐训练：目标限位 + 力矩裁剪）
        target = DEFAULT_ANGLES + action * ACTION_SCALE
        # 关节目标限位（与训练一致）
        jnt_lo = model.jnt_range[:, 0].copy() if hasattr(model, 'jnt_range') else None
        jnt_hi = model.jnt_range[:, 1].copy() if hasattr(model, 'jnt_range') else None
        # MuJoCo model.actuator_trnid 可能不直接暴露，改用 model.jnt_range
        try:
            trnid = model.actuator_trnid[:, 0]  # transmission joint indices
            lo = model.jnt_range[trnid, 0]
            hi = model.jnt_range[trnid, 1]
        except Exception:
            lo = np.full(12, -12.0)
            hi = np.full(12, 12.0)
        target = np.clip(target, lo, hi)
        torques = KP * (target - data.qpos[7:19]) - KD * data.qvel[6:18]
        data.ctrl[:] = np.clip(torques, -CLIP_TORQUES, CLIP_TORQUES)
        mujoco.mj_step(model, data)
        view.sync()

        # Time sync (policy at 50Hz = 0.02s per step)
        expected = step * 0.02
        elapsed = time.time() - loop_t0
        if elapsed < expected:
            time.sleep(expected - elapsed)

        step += 1

    kb.stop(); view.close()


if __name__ == "__main__":
    main()