jermo1120/b1k-pi05-eval

BEHAVIOR-1K 2026 Challenge — π0.5 policy evaluation

Evaluation of an openpi π0.5 policy on BEHAVIOR-1K 2026 Challenge Task 22 = putting_shoes_on_rack (robot b1k/R1Pro), plus setup notes for an abandoned Task 38 (spraying_for_bugs) LoRA training run.

Policy under test

Model	openpi π0.5, train config pi05_b1k, action_horizon=32, pi05=True
Base	gs://openpi-assets/checkpoints/pi05_base
Fine-tune	50,000 steps on behavior-1k/2026-challenge-demos (LeRobot v3), prompt = "putting shoes on rack"
Checkpoint	shoes_on_rack_v2_49999 (not included here — 21 GB)

How the eval works

• Policy served over a websocket by openpi/scripts/b1k/serve_b1k.py (JAX, one L40S).
• omnigibson.eval.eval drives the OmniGibson Evaluator: for each test instance it resets the sim and runs a rollout, querying the policy server for actions.
• Split public_test = 20 instances (ids 301–320). CPU dynamics (the official challenge config).
• Per-rollout metric q_score ∈ [0,1] (partial credit for goal conditions). Success = q_score == 1; task success rate = fraction of instances with q_score == 1.

Result

Task success rate = 0 / 20 = 0.0%. Mean q_score 0.015, best instance 0.2 (id 314). Every episode ran to the full 11,590-step timeout.

Failure mode: normalized_agent_distance.base (base path length ÷ human-demo base path length) is >1 — mostly 2–4× — on nearly every instance, while arm motion sits at ~0.2–0.8 (under-moving). The policy drives the base around far more than the demos and barely manipulates anything. No real task progress on any instance.

Full per-instance table: eval/EVAL_SUMMARY.md.

id	q_score	nd_base	nd_left	nd_right
301	0.000	2.99	0.34	0.34
302	0.000	2.05	0.42	0.50
303	0.000	1.58	0.36	0.42
304	0.000	2.66	0.33	0.33
305	0.100	1.08	0.21	0.27
306	0.000	3.48	0.56	0.59
307	0.000	2.26	0.43	0.49
308	0.000	1.79	0.38	0.42
309	0.000	3.02	0.51	0.49
310	0.000	2.37	0.39	0.49
311	0.000	2.19	0.42	0.37
312	0.000	2.47	0.42	0.42
313	0.000	0.72	0.14	0.16
314	0.200	2.55	0.33	0.38
315	0.000	2.84	0.43	0.45
316	0.000	4.39	0.84	0.86
317	0.000	1.80	0.29	0.31
318	0.000	3.56	0.68	0.67
319	0.000	2.22	0.26	0.26
320	0.000	3.39	0.67	0.69

Repository layout

eval/
  EVAL_SUMMARY.md            full status, run commands, all fixes applied during bring-up
  NOTES.md                   original setup notes (disk-quota bring-up)
  TASK22_EVALUATION.md       task-22 eval design notes
  run_task22_eval.sh         the eval orchestrator (policy server + eval client, detached)
  env2.sh                    environment file (installs redirected off the quota'd volume)
  eval_task38_at_10k.sh      unused: planned mid-training eval for task 38
  build_task38_manual.py     builds the standalone task-38 LeRobot dataset
  *.log                      setup / eval / training logs
  results/json/*.json        20 per-instance result JSONs
  results/videos/*.mp4       20 rollout videos (Git LFS)
  frames/                    sampled frames from instance 301
code-patches/
  openpi_behavior.patch      diff vs wensi-ai/openpi@behavior (task + robot-name + task-38 configs)
  behavior1k_v3.9.2.patch    diff vs BEHAVIOR-1K v3.9.2 (eval.py / obs_utils.py)
  *_base_commit.txt          the upstream commit each patch applies onto
notebooks/
  view_task22_314.ipynb      plays the best rollout inline

Rollout videos (Git LFS)

The 20 MP4s under eval/results/videos/ are tracked with Git LFS. To fetch them:

git lfs install
git clone <repo-url>
git lfs pull

Task 38 (spraying_for_bugs) — abandoned

A π0.5 LoRA training run on task 38 was started concurrently but crashed at ~step 20k/50k when the orbax checkpoint write hit the RunPod volume's disk quota (checkpoint_base_dir resolved onto the quota-capped path). No checkpoint was ever saved and the run was not restarted. Dataset build script and config patches are included for reference only.