davidwdw/b1k

GPT-6 generated BEHAVIOR-1K policies — Tasks 00, 01, 02 & 63

The original Task 01/02 policies were collected from eai on 2026-09-14. They are standalone code (no Codex / LLM / API calls at run time), evaluated with the official BEHAVIOR v3.9.2 evaluator. Later Task 00 and Task 63 collections are documented below; Task 63 uses a local Qwen model.

Task	Source on eai	Run set	Runs	Successes	Mean final Q
01 picking_up_trash	/home/eai/yixin/gpt6_b1k	final_prior_frozen20 (instances 301–320, seed 0)	20	13	0.767
01	rerun on this laptop (2026-09-14)	rerun_local_20260914 (same frozen code, + videos) — REPORT	20	14	0.800
02 putting_away_Halloween_decorations	/home/eai/yixin/code/gpt6cap	instances301-320_seed0 (skip 312)	19	0	0.293
02		instance312_seeds0-4 (seed 5 interrupted, no result)	5	0	0.657
02		repeat_20260912_6inst (301, 303, 308, 317, 318, 320)	6	0	0.238

Per-run numbers: scores.csv.

Layout

gpt6_b1k/
  README.md  scores.csv
  SHA256SUMS                 # checksum of every file below (sha256sum -c SHA256SUMS)
  MANIFEST.tsv               # path, sha256, original eai path (or how a file was derived)

  task01_picking_up_trash/
    code/                    # exact frozen snapshot that was evaluated (identical to repo head)
                             #   gpt6_b1k/ policy, config/, assets/ (models + scene prior), third_party/,
                             #   run.py, env.sh, tests/, README/EVALUATION/EXPERIMENTS/PRIOR_DESIGN.md
    final_prior_frozen20/
      summary.json manifest.json verification.json remote_verification.json launcher_final.py logs/
      instance_<id>/
        result.json          # official evaluator output
        traj/actions.jsonl   # per-step actions (split from batch file)
        traj/events.jsonl    # policy events (split from batch file)
        traj/observations/<episode>_<step>/   # sampled RGB + linear depth (.npy) for 3 cameras + observation.json
      instance_304/interrupted_attempt_batch000/traj/   # 304 attempt killed by transport failure (no score)
      raw_batches/batch_000, batch_001/   # original unsplit actions/events, evaluator/policy logs, command.json
    rerun_local_20260914/    # rerun on cesar-ThinkPad-P1-Gen-7 (RTX 3000 Ada), identical code/evaluator, video on
      REPORT.md summary.json manifest.json logs/{launcher.log,vram.csv} source/ raw_batches/batch_000/
      instance_<id>/ result.json  video.mp4  traj/{actions,events}.jsonl  traj/observations/
    earlier_runs/            # development runs in their original layout (partial; contain the only Task 01 videos)
      prior_frozen20_18/ visual_frozen20_16/ validation_pack09/ validation_floor10/

  task02_putting_away_halloween_decorations/
    code/                    # r21 policy: source/ (pipeline + scripts), reference/, launch_policy.py,
                             #   evaluate_seed.py, run_centre*.sh, run.sbatch, suites, batch_tools/, batches/,
                             #   checksum manifests, README/CODE_LAYOUT/BATCH_RUNS.md
    instances301-320_seed0/  # + RESULTS.md, FAILURE_ANALYSIS.md, FINAL_REPORT.md, summary.json, previews/
      instance_<id>/ result.json  video.mp4  traj/{actions,events}.jsonl  logs/{evaluator,policy}.log exit_code.txt
    instance312_seeds0-4/seed_<nn>/     # same per-run layout; seed_05 has logs/traj only
    repeat_20260912_6inst/instance_<id>/seed_000/attempt_001/   # same per-run layout (+ command.json, run_status.json)
                             # batch-level code/ snapshot, manifest.json, REPORT.md, FINAL_REPORT.md

Notes

• Task 01 local rerun (rerun_local_20260914): same 673 frozen files and evaluator hashes as the eai run, re-run here with --write-video; 19/20 instances match eai's Q exactly, 318 went 0.333 → 1.0 (different can order). Launcher: ~/code/gpt6_b1k/runs/launch_frozen_local.sh; files added later are appended to SHA256SUMS/MANIFEST.tsv by organize_rerun.py.
• Task 01 eai final run has no videos. Video encoding was off (official default). Only sampled RGB-D observations exist; the 8 Task 01 videos are in earlier_runs/*/batch_000/evaluation/videos/.
• Task 01 trajectory split. The evaluator ran all instances in one process per batch, so the policy wrote one actions.jsonl/events.jsonl per batch. Even episode numbers map to instances in order (batch_000: 2→301, 4→302, 6→303, 8–10→interrupted 304; batch_001: 2→304 … 34→320). Odd episodes are reset-only and were assigned to the following instance. Line counts of the split files sum exactly to the originals, which remain in raw_batches/.
• Task 02 raw sensor observations were not collected. They exist only on Centre (/home/david/r21-seed10/runs/...).
• Original Markdown reports were copied unchanged. Their links point to absolute eai paths / old folder names; use MANIFEST.tsv to map them.
• Renamed per-run files: evaluation/json/*.json → result.json, evaluation/videos/*.mp4 → video.mp4, policy/*.jsonl → traj/, logs → logs/.
• Not copied: gpt6_b1k debug/ (18 GB dev lab sessions), appdata/ (9.2 GB Isaac cache), observations of Task 01 earlier runs (~4.5 GB), gpt6cap official/ (565 MB unmodified BEHAVIOR v3.9.2 export; its checksums are in task02_.../code/OFFICIAL_SHA256SUMS), __pycache__, .sync.lock.
• index.html is the results viewer (regenerate with python3 build_index.py).
• Integrity: every copied file was verified against SHA256 computed on eai before reorganizing; SHA256SUMS verified again after.

Keeping this folder in sync — sync/

From 2026-09-18 on, remote B1K code and results are pulled with python3 sync/sync_centre.py. The script can be re-run safely. It mirrors each registered source, verifies every file against SHA-256 computed on the remote machine, and records history in sync/SYNC_LOG.jsonl. See sync/README.md for what is copied, what is skipped, and how to add a new source.

Task 01 · Centre human-first continuation teacher

task01_picking_up_trash/centre_human_first_recovery/ mirrors centre:/media/david/HDD/gpt6_b1k/task01_human_first_recovery. It is a bimanual teacher that carries the bin and can continue from any human-demo frame, a failed VLA state, or the official initial state. The mirror holds code (code/, code_v2/ git repos), per-run frozen snapshots, reports and admission lists, and every run's outcome, audits, and video. It includes the trajectory.hdf5 of the training-eligible runs only (34 files, ~117 GB); the other ~0.97 TB of HDF5 stays on Centre and is listed in _sync/REMOTE_INVENTORY.tsv. Per-batch results are in _sync/RUNS_SUMMARY.tsv. Human-start continuation reaches 14/16 in human-random20-v7; from the official initial state, 8, 7, and 9 of 20 on public test 301–320. These files are not added to the root SHA256SUMS / MANIFEST.tsv; they have their own _sync/SHA256SUMS.

Task 00 collection — 2026-09-18

Task 00 · turning_on_radio now contains the full 4080 production export: 300 instances (259 successes, 41 failures), both original LeRobot v3 datasets, frozen source and matching official evaluator, all original attempts and audit logs, and the Chinese distribution report. Open its results index. All copied files were verified against source SHA256. The generated depth videos retain the documented lossy encoding difference from official human demonstrations. Original Task 01/02 files are unchanged.

Within Task 00, code/ holds source; aligned_20260918/ holds per-instance results; datasets/task00_all_v1/ and datasets/task00_success_v1/ preserve their training dataset structures. Task 00 has its own summary.json and HTML index; the older root scores.csv remains the Task 01/02 score table.

Task 01 recovery datasets — 2026-09-21

The full completed recovery campaign and its filtered primitive subset are now mirrored under task01_picking_up_trash/, including exact generation code, HDF5 trajectories, and per-source SHA-256 verification. See local dataset paths and loading instructions. Registered refresh commands are in sync/README.md.

Task 63 collection — 2026-09-22

Task 63 · installing_smoke_detectors includes the exact candidate18 policy, evaluator, local Qwen3.5-0.8B checkpoint, and retained public evaluation records: 20/20 successes on instances 301–320. This configuration used a local language model; it requires no external API.

The training datasets separately cover 300 training instances: 271 successes and 29 failures. task63_all_v1 has 300 episodes / 829,800 frames; task63_success_v1 has 271 episodes / 718,034 frames. Their original LeRobot v3 layouts and checksum manifests are preserved.

Synchronization scope

The consolidated collection is maintained locally, in the Hugging Face dataset davidwdw/b1k, and in centre:/media/david/HDD/gpt6_b1k_collection/. Centre's live source directories remain separate. See sync/README.md for transfer directions and verification. Syncs preserve destination-only files, exclude active campaign runs, and retain the existing historical HDF5 eligibility rule. Simulator caches, local assistant state, upload caches, and nested Git internals are not part of the published payload.