DM0.5 LIBERO for LeRobot

Mirror of Dexmal/DM05-Lerobot-LIBERO 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_libero is a LIBERO-specific fine-tuned checkpoint based on lerobot/dm05_base. It is not the general-purpose DM05 base model.

Model Configuration

Cameras: front, wrist
State dimension: 8
Action dimension: 7
chunk_size=10
n_action_steps=10
add_state=false
Processor: use_relative_actions=false

For evaluation, use the LIBERO relative-controller protocol. The processor keeps the model action representation absolute; the controller protocol is an evaluation-time environment setting.

LIBERO Evaluation

The reported 50k-step result covers 40 tasks x 5 episodes per task = 200 episodes:

Suite	Successes
Spatial	49/50
Object	50/50
Goal	50/50
LIBERO-10	48/50
Total	197/200 = 98.5%

This is a 200-episode evaluation result, not a standard 2,000-episode complete LIBERO evaluation score.

MUJOCO_GL=egl lerobot-eval \
  --policy.path=lerobot/dm05_libero \
  --env.type=libero \
  --env.task=libero_spatial,libero_object,libero_goal,libero_10 \
  --env.camera_name_mapping='{"agentview_image":"front","robot0_eye_in_hand_image":"wrist"}' \
  --env.observation_height=256 \
  --env.observation_width=256 \
  --env.control_mode=relative \
  --eval.n_episodes=5 \
  --eval.batch_size=1 \
  --seed=7 \
  --policy.device=cuda

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}
}