基于PI05,在ROBOTWIN 2.0的一个小任务中微调，使用RISE方法的offline training代码，微调2000STEP后在ROBOTWIN上仿真效果全为0

Start with the posted TrainConfig for PI05, RISE offline training, and the ROBOTWIN Hanging_Mug task. Check the model, action dimensions, data transforms, normalization assets, and training schedule against the reported zero-result simulation after 2,000 steps. Done means identifying the compatibility mismatch and confirming nonzero ROBOTWIN simulation results.

VLA新手可能不太了解各个模型，benchmark和方法之前的兼容性。
TrainConfig(
name="Policy_offline_Hanging_Mug_robotwin_compatible",

    model=pi0_config.Pi0Config_Custom(
        pi05=True,
        action_dim=32,
        valid_action_dim=14,             # ← 关键修复
        action_horizon=50,               # Pi0Config 默认是50，保持一致
        freeze_vlm_backbone=True,
    ),

    data=LeRobotAlohaDataConfig(
        repo_id="Hanging_Mug",

        adapt_to_pi=False,

        assets=AssetsConfig(
            assets_dir="../../datasets/norms",
            asset_id="hanging_mug_robotwin_compatible_stats",
        ),

        repack_transforms=_transforms.Group(
            inputs=[
                _transforms.RepackTransform(
                    {
                        "images": {
                            "cam_high": "observation.images.cam_high",
                            "cam_left_wrist": "observation.images.cam_left_wrist",
                            "cam_right_wrist": "observation.images.cam_right_wrist",
                        },
                        "state": "observation.state",
                        "actions": "action",
                        "prompt": "task",
                    }
                )
            ]
        ),

        base_config=DataConfig(
            prompt_from_task=True,
        ),
    ),

    lr_schedule=_optimizer.CosineDecaySchedule( 
        warmup_steps=200,      # ← 和5000步匹配，warmup占4%
        peak_lr=5e-5,          # ← batch大了适当调大
        decay_steps=5000,      # ← 和 num_train_steps 对齐
        decay_lr=5e-6,
        ),
    
    pytorch_weight_path="../pretrained_openpi/pi05_base/",
    
    num_train_steps=5000,
    save_interval=2000,
    keep_period=1000,
    batch_size=128,
    num_workers=8,
),