The adapter retains the OpenDM model path while using LeRobot’s standard datasets, processors, training, checkpointing, Hub loading, and evaluation flow.
The page specifies lerobot/dm05_base as a self-contained conversion of the raw OpenDM checkpoint and says it can be loaded with DM05Policy.from_pretrained(). It accepts standard LeRobot datasets with image or video observations, observation.state, actions, and task descriptions. The documented LIBERO recipe uses 50,000 training steps, batch size 8, action chunks of 10, and CUDA. A provided lerobot/dm05_libero checkpoint recorded 197 successes in 200 evaluation episodes across four LIBERO suites.
That result is explicitly a 200-episode reproduction check, not the 2,000-episode LIBERO benchmark protocol. The documentation also says that a complete checkpoint directory is required; a standalone model.safetensors file is insufficient. The page is for LeRobot’s main version, which requires installation from source, so it does not establish that DM05 support is available in the latest stable pip release.