Teleop-Fix: data, weights and agent transcripts

Companion data for https://github.com/NapoleonPK/teleop-fix: a pre-registered simulation study of whether an AI agent (Claude Code, claude-opus-5-5) can diagnose and fix a behaviour-cloning robot policy trained on flawed teleoperation demonstrations (hesitant operator, leader/follower recording lag, workspace comfort zone, clean controls), compared with recording more demonstrations, a heuristic filter, longer training and random data search. Results, code and the frozen pre-registration are in the GitHub repository.

Contents

• units/u<N>/: one folder per unit (test u1-u12, dev u51-u52)
  • dataset.npz: 500 demonstrations, X observations (500 x 100 x 31), A actions (500 x 100 x 4)
  • policy.npz: the broken (or clean) ACT-lite policy trained on it
  • failed_episodes/: the 3 failed episodes shown to the agent
  • arms/: baseline weights (demos_10/25/50, filter, longer, search) and log.json
  • meta.json: operator profile and seeds
• sessions/<id>/: every agent session (12 test, 2 dev)
  • transcript.jsonl: full Claude Code stream
  • work/: the agent's scripts and files; work/submission/ holds policy.npz and report.md
  • manifest.json, audit.json, service.jsonl (rollout and training log)
• results/: eval_heldout.json (held-out success, all arms), diagnosis_scores.json and diagnosis_reasons.json, fix_kinds.json, sessions.json, teleopfix.md and teleopfix.json (analysis output)
• superseded/: kept for transparency, not used in the results
  • failed_auth/: two sessions that ended on a login error after one turn
  • failed_usage_limit/: sessions cut by the account usage limit, then rerun
  • rebuilt_desync_padding/, rebuilt_desync_k2/: desync units replaced before the freeze (see DECISIONS.md 8 and 10)

Weights are .npz state dicts readable by tf/policy.py (torch) or tf/policy_np.py (numpy) in the repository.

License: CC BY 4.0.