k1000dai/vla-agent-loop-runs

vla-agent-loop runs

Raw artifacts from vla-agent-loop self-improvement runs on an SO-101 arm (Jul 30 – Aug 1, 2026): real-robot rollouts and the pi05 SFT checkpoints trained on them.

Layout

run-<date>-<time>/
  state.json, history.jsonl        loop state
  artifacts/
    rollouts/collect_<N>/
      metadata.yaml
      episodes/<ep>.tar            one uncompressed tar per episode
    train/vla_train_pi05_sft_*/<step>/
      pretrained_model/            LeRobot policy (model.safetensors, configs, normalizers)
      training_state/              optimizer / scheduler / rng state
    dataflows/, logs/, pi05/, hpc/ dora dataflows, logs, bookkeeping

Each episodes/<ep>.tar extracts to <ep>/ with:

• cameras/{top,side,side2,wrist}/<frame>.jpeg
• obs/arms/follower/qpos.parquet
• action/arms/follower/qpos.parquet

MANIFEST.jsonl lists every uploaded file with its sha256 and size (and member count for tars). SYMLINKS.json lists the last -> <step> checkpoint symlinks that existed locally.