Backup DreamWaQ rslrl stability fixes

scripts/dreamwaq_sim2sim_mujoco.py

@@ -31,7 +31,8 @@ HISTORY_LEN = 5
 ACTION_SCALE = 0.25
 KP = 28.0
 KD = 0.7
-CLIP_ACTIONS = 23.7
+CLIP_ACTIONS = 4.0
+CLIP_TORQUES = 80.0
 CLIP_OBS = 100.0
 MAX_VX, MAX_VY, MAX_WZ = 1.0, 1.0, 1.0
 
@@ -122,7 +123,7 @@ def main():
 
     # Select XML scene
     terrain_map = {
-        "flat": "scene_motor_actuator.xml",
+        "flat": "scene_dreamwaq_flat.xml",
         "rough": "scene_rough_terrain.xml",
         "stairs": "scene_stairs_terrain.xml",
         "dreamwaq": "scene_dreamwaq_terrain.xml",
@@ -243,10 +244,22 @@ def main():
             action = np.clip(action, -CLIP_ACTIONS, CLIP_ACTIONS)
             last_action = action.copy()
 
-        # PD control
+        # 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_ACTIONS, CLIP_ACTIONS)
+        data.ctrl[:] = np.clip(torques, -CLIP_TORQUES, CLIP_TORQUES)
         mujoco.mj_step(model, data)
         view.sync()