eliasab16/pi05_recap_batch_1_bc

pi0.5 BC fine-tune on SO-110 wire-insertion

Flow-matching behavior cloning of lerobot/pi05_base on 130 episodes of SO-110 teleoperated wire-insertion (eliasab16/pi05_recap_batch_1).

Training recipe

• Policy: pi05 (flow-matching only, no FAST tokens)
• Action encoding: relative actions (use_relative_actions=true), gripper excluded
• Action chunk: 50 steps
• Batch size: 16 (single-GPU)
• Steps: 10,000 (~1.45 epochs)
• dtype: bfloat16
• Grad checkpointing: yes
• Vision encoder: trained (not frozen)
• Hardware: 1× A100 80GB
• Final loss: 0.014
• Wall time: ~8h 23m at 2.81 s/step

Checkpoint contents

• pretrained_model/ — model.safetensors (the policy weights) + config + preprocessor/postprocessor
• training_state/ — optimizer / scheduler / RNG state for resuming

Important: relative-action stats

The default eliasab16/pi05_recap_batch_1 dataset on the Hub has stats computed from absolute actions. To resume training (or re-run inference with correct normalization), use the relative-action stats.json shipped in this repo under dataset_stats/meta/stats.json. Copy it over your dataset's meta/stats.json, or regenerate with:

lerobot-edit-dataset \
  --repo_id=eliasab16/pi05_recap_batch_1 \
  --operation.type=recompute_stats \
  --operation.relative_action=true \
  --operation.chunk_size=50 \
  --operation.relative_exclude_joints='["gripper"]' \
  --operation.overwrite=true

Resume training

lerobot-train \
  --dataset.repo_id=eliasab16/pi05_recap_batch_1 \
  --policy.type=pi05 \
  --policy.pretrained_path=eliasab16/pi05_recap_batch_1_bc \
  --policy.use_relative_actions=true \
  --policy.relative_exclude_joints='["gripper"]' \
  --policy.dtype=bfloat16 \
  --policy.gradient_checkpointing=true \
  --policy.device=cuda \
  --batch_size=16 \
  --num_workers=2 \
  --save_freq=500 \
  --steps=20000 \
  --output_dir=./outputs/pi05_recap_resumed \
  --resume=true