Long-WAM LIBERO IDM — LeRobot

LeRobot-native version of Long-WAM LIBERO IDM. IDM imagines future video before denoising actions. This checkpoint uses 48 control ticks of observation history.

Use a LeRobot version with Long-WAM policy support and the longwam and libero extras installed. The repository includes the complete policy weights, configuration, and preprocessing statistics. The tokenizer loads from google/umt5-xxl.

Evaluation

lerobot-eval \
  --policy.path=Efficient-Large-Model/Long-WAM-LIBERO-IDM-LeRobot \
  --policy.device=cuda \
  --env.type=libero --env.task=libero_spatial --eval.batch_size=1

For direct inference, load LongWAMPolicy.from_pretrained(...) with the saved pre/post-processors. Call select_action on every control tick and reset() between episodes. Use the saved camera order and action/state definitions. Execution is synchronous.