fix: dt in state.reward, feet_air_time first_contact timing

motrix_rl/src/motrix_rl/tasks/go1.py

@@ -105,7 +105,7 @@ class rslrl:
     class Go1DreamWaQWalkRslrlPpo(RslrlCfg):
         """Go1 DreamWaQ walk — CENet VAE + 不对称特权观测。"""
 
-        num_envs: int = 1024  # 上游 4096，CPU/GPU 安全默认
+        num_envs: int = 2048
 
         def __post_init__(self):
             runner = self.runner
@@ -126,6 +126,7 @@ class rslrl:
             runner.algorithm.entropy_coef = 0.01      # 上游 Go1RoughCfgPPO
             runner.algorithm.desired_kl = 0.01        # 上游 0.01 (默认 0.008)
             runner.algorithm.clip_param = 0.2
+            runner.algorithm.schedule = "fixed"   # 固定 schedule，防止 noise_std 发散
             runner.algorithm.gamma = 0.99
             runner.algorithm.lam = 0.95
             runner.algorithm.max_grad_norm = 1.0
@@ -135,7 +136,7 @@ class rslrl:
             runner.actor.class_name = (
                 "motrix_rl.rslrl.torch.models.cenet_actor:CENetActorModel")
             runner.actor.hidden_dims = [512, 256, 128]
-            runner.actor.init_noise_std = 1.0  # 上游原值
+            runner.actor.init_noise_std = 0.5  # 降低探索噪声，1024 envs 不需要太高
 
             # Critic：标准 MLPModel，输入 privileged_obs
             runner.critic.class_name = "MLPModel"