cyy/Motrixlab
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: MuJoCo sim2sim default_angles match MotrixSim training

Browse Source
This commit is contained in:
8x54zj-m
2026-07-01 20:55:16 +08:00
parent 2bd4c05e1b
commit 63b235d186
3 changed files with 8 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
motrix_rl/src/motrix_rl/rslrl/torch/train/dreamwaq_ppo.py
View File

@@ -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；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))
# CV → bootstrap 概率（CENet vel head 通过 VAE loss 已有梯度，无需额外保护）
self.actor._adaboot_prob = max(0.0, min(1.0, cv * 5.0))
def _compute_surrogate_loss(
self, actions_batch, actions_log_prob_batch,

2
motrix_rl/src/motrix_rl/tasks/go1.py
View File

@@ -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 = 0.5 # 与 fixed schedule 配合
runner.actor.init_noise_std = 1.0 # 论文原值
# Critic：标准 MLPModel，输入 privileged_obs
runner.critic.class_name = "MLPModel"