wangbin08/DM05-BiGym-CutleryLong-50pct

DM05 BiGym DishwasherUnloadCutleryLong 50%

Standalone BF16 OpenDM DM0.5 checkpoint for BiGym DishwasherUnloadCutleryLong. The task LoRA is merged into the Apache-2.0 Dexmal/DM05 base weights, so a separate base-model download is not required. This is the checkpoint used as the 50% native baseline in the AMD 3DGS domain-adaptation experiments.

Results

Visual domain	Success rate	Successful episodes
Native BiGym	50%	25/50
3DGS kitchen shell d7b54c6a	40%	20/50

Protocol: seeds 200000-200049, 20 Hz control and policy, chunk size 16, 10 diffusion steps, execute 8 actions per inference call, maximum 800 policy steps, model sequence length 896, and prompt Unload cutlery from dishwasher to drawer task..

Training

• Base model: Dexmal/DM05
• Dataset: 32 successful DishwasherUnloadCutleryLong trajectories from the referenced BiGym Plus revision
• Observations: head and two wrist RGB cameras plus 16D state
• Action: 16D absolute joint-position command, including cumulative floating base, arm joints, and grippers
• LoRA: rank 32, alpha 16, all linear modules
• Trainable parameters: 324,281,376 of 6,153,393,584 (5.27%)
• Hardware: four RTX 5090D GPUs
• Per-device batch: 12; global batch: 48
• Steps: 10,000; model max length: 896

Usage

Use this model directly with the pinned OpenDM integration code:

git clone https://github.com/Kyrie-w8/amd-bigym-3dgs-opendm.git

The repository contains the native and 3DGS evaluators, the exact fixed-seed protocol, and the reproducible LoRA-to-full-model merge utility.

Limitations

• This checkpoint is task-specific and is not a general-purpose robot policy.
• 50% refers to native BiGym only; the measured 3DGS-shell result is 40%.
• Results are simulator-, seed-, camera-, action-definition-, and protocol-specific.
• The 3DGS source scene is governed by separate DL3DV terms and is not included.