Dexmal/DM05-Lerobot-MEM

DM05-MEM for LeRobot

This is a LeRobot-format conversion of the public Dexmal/DM05-MEM base weights, adapted for LeRobot datasets, processors, training, and inference. The BF16 weights are unchanged. Load the complete checkpoint with --policy.path or DM05Policy.from_pretrained(); no checkpoint conversion is required.

Requirements

Use Python 3.12 or later and a LeRobot checkout with DM05-MEM support. From the LeRobot repository root, install the training dependencies:

pip install -e '.[training,dataset,dm05]'

Fine-tuning

Prepare a standard LeRobot dataset containing images, observation.state, action, and task descriptions, then fine-tune directly:

lerobot-train \
  --policy.path=Dexmal/DM05-Lerobot-MEM \
  --dataset.repo_id=your_org/your_dataset \
  --policy.use_memory=true \
  --policy.memory_image_key=observation.images.cam_high \
  --policy.memory_frames=20 --policy.memory_stride=25 \
  --policy.chunk_size=25 --policy.n_action_steps=25 \
  --policy.push_to_hub=false \
  --policy.device=cuda \
  --batch_size=1 \
  --steps=1000 \
  --output_dir=outputs/train/dm05-mem

For a local copy, replace the Hub identifier with the checkpoint directory. Training uses the target dataset's normalization statistics. Match its camera names to the checkpoint with --rename_map, for example:

--rename_map='{"observation.images.top":"observation.images.cam_high","observation.images.left_wrist":"observation.images.cam_left_wrist","observation.images.right_wrist":"observation.images.cam_right_wrist"}'

Use policy.image_keys to set camera order and policy.memory_image_key to select the history camera. Adjust the history stride to the dataset frame rate; memory_stride=25 samples one past frame per second for 25 FPS data. Set policy.robot_type for the target robot prompt and use state/action statistics appropriate to the target dataset and action representation. Fresh fine-tuning uses the target dataset's feature dimensions and statistics; inference uses those saved with the fine-tuned checkpoint.

Default Configuration

The checkpoint includes the following configurable LeRobot defaults:

Setting	Value
State and output action	14-dimensional absolute joint values
Cameras, in order	cam_high, cam_left_wrist, cam_right_wrist
Image feature prefix	observation.images.
Main memory camera	observation.images.cam_high
Past frames / stride	20 / 25 dataset or captured control frames
Input capture rate for this profile	25 FPS
Prediction / execution chunk	25 / 25
Robot prompt	Aloha
Normalization	Per-dimension quantiles; clipping disabled

These are adapter defaults, not restrictions on the fine-tuning task. The saved joint-space processor profile differs from the source release's native EEF inference profile. Use target-dataset statistics for fine-tuning rather than reusing the bundled statistics for a different robot or action representation.

Loading and History

from lerobot.policies.dm05.modeling_dm05 import DM05Policy
from lerobot.policies.factory import make_pre_post_processors

checkpoint = "Dexmal/DM05-Lerobot-MEM"
policy = DM05Policy.from_pretrained(checkpoint)
preprocessor, postprocessor = make_pre_post_processors(
    policy.config, pretrained_path=checkpoint,
)

Feed every control-frame observation through the saved preprocessor and select_action(), then use the saved postprocessor. Reset the policy and processors between episodes. Direct predict_action_chunk() calls require an explicit history sequence and padding masks, not one current image per action chunk.

Checkpoint Format

The directory contains config.json with nested core_config, BF16 model.safetensors, pre/postprocessor configurations and normalization tensors, and the tokenizer/image-processor assets in dm05_processor/. conversion.json records the conversion provenance and saved processor profile. The redundant tied input-embedding alias is omitted from the saved state dict; its values are reconstructed from the tied output embedding on load.

Resources

• Original DM05-MEM model
• OpenDM repository
• LeRobot repository
• DM0.5 technical blog

License

The source model is distributed under the Gemma license. This conversion retains that license and its applicable terms; it does not grant additional rights. See the source model card.