Align DreamWaQ PPO rollout and policy dynamics

2026-07-22 13:59:16 +08:00
parent 57faf8880b
commit 0555776c11
7 changed files with 26 additions and 15 deletions
--- a/motrix_envs/src/motrix_envs/locomotion/go1/dreamwaq.py
+++ b/motrix_envs/src/motrix_envs/locomotion/go1/dreamwaq.py
@@ -197,7 +197,7 @@ class DreamWaQTask(Go1WalkTask):

    def apply_action(self, actions, state):
        """裁剪动作 + 随机延迟（模拟真实部署延迟，提高 sim-to-real 泛化）。"""
-        actions = np.clip(actions, -4.0, 4.0)
+        actions = np.clip(actions, -100.0, 100.0)
        # 随机动作延迟：0-3 个控制步（0-60ms）
        if not hasattr(self, '_action_buffer'):
            self._action_buffer = np.zeros((self._num_envs, 3, self._num_action), dtype=np.float32)