DM0.5 for LeRobot

Mirror of Dexmal/DM05-Lerobot by Dexmal, with the processor pipelines regenerated for the DM05 policy as merged in LeRobot. Weights, config.json and normalization statistics are unchanged from the original.

lerobot/dm05_base is the LeRobot-format base checkpoint of DM0.5, adapted from OpenDM. It predicts continuous action chunks from images, robot state, and language instructions.

This is a base model for supervised fine-tuning, not a LIBERO-, RoboTwin-, or robot-specific checkpoint.

Fine-tuning

lerobot-train \
  --dataset.repo_id=HuggingFaceVLA/libero \
  --policy.path=lerobot/dm05_base \
  --policy.add_state=false \
  --policy.chunk_size=10 \
  --policy.n_action_steps=10 \
  --policy.repo_id=your_repo_id \
  --output_dir=outputs/train/dm05-libero \
  --steps=50000 \
  --batch_size=8 \
  --policy.device=cuda

For local-only training, replace --policy.repo_id=... with --policy.push_to_hub=false.

This LIBERO recipe matches OpenDM: it excludes state from the prompt and learns stored actions unchanged. Keep policy.add_state=true unless the target recipe specifies otherwise. Policy action representation and environment control mode are configured independently.

Training uses the standard LeRobot state/action statistics in meta/stats.json. With policy.use_relative_actions=true the arm targets become action - state while the statistics stay the dataset's own, so the normalization scale is fitted to absolute actions. This trains, with somewhat less signal, and is a general LeRobot limitation of relative actions rather than a DM05-specific one.

Evaluation

Evaluate a fine-tuned checkpoint:

MUJOCO_GL=egl lerobot-eval \
  --policy.path=/path/to/checkpoint/pretrained_model \
  --env.type=libero \
  --env.task=libero_spatial \
  --env.control_mode=relative \
  --policy.device=cuda

Checkpoint contract

The base checkpoint uses OpenDM's 14-dimensional state/action schema. Fresh SFT takes its effective feature schema and statistics from the target LeRobot dataset.

Load the complete checkpoint directory with DM05Policy.from_pretrained() or --policy.path; policy config, tokenizer, preprocessing state, and weights are all required.

Resources

Citation

@misc{dm05,
    title  = {{DM0.5}: An Open-World Foundation Model for General-Purpose Embodied Intelligence},
    author = {{Dexmal Team}},
    month  = {July},
    year   = {2026},
    url    = {https://www.dexmal.com/blog/dm0.5/index_en.html}
}