pi0_5_robodojo_taco_hold_b64_60k

Low-level policy for RoboDojo long-horizon (8 real-robot bimanual tabletop tasks, 100 demonstrations each), fine-tuned from lerobot/pi05_base on Myungkyu/RoboDojo-taco-gemini — demonstrations with dense subtask labels from the task-specific context (offline annotation).

• Architecture: Pi0.5 vanilla (three live camera views, no keyframe slot) trained with subtask-hold action chunks: every training chunk is cut where the current subtask ends and the remaining steps repeat the last in-subtask target, so the policy stops instead of carrying on with the demonstration's next subtask
• Optimizer batch 64, 60000 steps, final checkpoint
• Inputs: head + left/right wrist images, proprioception, the current subtask text (no keyframe input)
• Subtask hold: action_hold_after_subtask: true in config.json (lerobot fork field, ignored by stock lerobot); at inference re-plan as soon as the predicted chunk settles (the RoboDojo adapter's hold detector, LL_HOLD_DETECT=1)

Configs reference the base backbone / tokenizer by hub id or by the training site's local path — point them at your local copies before loading.