fix: fixed schedule, init_noise 0.5, AdaBoot cap 0.9 (keep vel head gradient)

motrix_rl/src/motrix_rl/rslrl/torch/train/dreamwaq_ppo.py

@@ -222,8 +222,8 @@ class DreamWaQPPO(PPO):
         cv = std_r / (mean_r + 1e-6)
         if not (0 <= cv < 1e6):  # NaN 或 Inf → 保持当前 prob
             return
-        # CV → bootstrap 概率（CV 高→不稳定→多用 GT；映射系数 5.0 可调）
-        self.actor._adaboot_prob = max(0.0, min(1.0, cv * 5.0))
+        # CV → bootstrap 概率（CV 高→不稳定→多用 GT；上限 0.9 保证 CENet vel head 有梯度）
+        self.actor._adaboot_prob = max(0.0, min(0.9, cv * 5.0))
 
     def _compute_surrogate_loss(
         self, actions_batch, actions_log_prob_batch,

motrix_rl/src/motrix_rl/tasks/go1.py

@@ -126,7 +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 = "adaptive"  # 上游原值，前期 bug 已修
+            runner.algorithm.schedule = "fixed"  # 4096 envs 采样够大，不需要 adaptive 加噪
             runner.algorithm.gamma = 0.99
             runner.algorithm.lam = 0.95
             runner.algorithm.max_grad_norm = 1.0
@@ -136,7 +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，adaptive 会自行调节
+            runner.actor.init_noise_std = 0.5  # 与 fixed schedule 配合
 
             # Critic：标准 MLPModel，输入 privileged_obs
             runner.critic.class_name = "MLPModel"