ethan-li-k/astra_bench/pi0-finetune

pi0_finetune — π₀ / π₀.₅ finetune on RoboRSI datasets

Overview

Finetune a pretrained π₀ / π₀.₅ checkpoint on a single-task LeRobot dataset produced by dataset/lerobot_build. No multi-task mixing in this skill — that belongs in a future training/pi0_multitask sibling.

When to use

• You have ≥ 50 successful episodes for a task.
• You want a fast-inference policy for this task that replaces

expert_replay or rollout_vlm in execution mode.

Phases

1. Validate dataset: LeRobotDataset(dataset) loads; camera set matches

the policy's expected input (else fail loud).

1. Resolve base checkpoint: physical-intelligence/pi0.5 or local path.
2. Launch lerobot-train subprocess with CLI args derived from params.
3. Stream logs; persist output_dir/training_state.json + final

checkpoint.

1. Return summary: final step, val loss (if split provided), output path.

Success criteria

• Final checkpoint directory non-empty.
• Val loss (if any) monotonic-ish (no divergence).
• Hook downstream evaluation/success_rate to verify actual robot

performance — loss is not a success proxy for manipulation.

Failure modes

Implementation status

SKELETON. The shipped policy.py prints the would-be lerobot-train command and exits. Wiring the actual subprocess requires deciding on the checkpoint cache location and pulling in HF auth / proxy handling from roborsi.config.loader. Lands in the next slice.

Why a skill and not a script

Because the task bundle needs to say "train then evaluate" declaratively without shell-script glue. Everything a task does is a skill.