task01_picking_up_trash/code/EVALUATION.md

Task 01: complete frozen-policy evaluation

All 20 official public instances (301–320) completed. Mean final Q = 0.766667; target Q > 0.7 achieved. 13/20 episodes have official success=true. Failed episodes remain in the mean. This is one known-scene, seed-0 evaluation, not an all-instance-success or unseen-scene claim.

Results

Each instance link opens its original official evaluator JSON. Machine-readable aggregation: summary. Verification checks all 673 frozen policy/asset/dependency files and the unmodified official evaluator files.

Instance	Final Q	Official success	Steps
301	1	yes	5,596
302	1	yes	6,123
303	1	yes	5,136
304	1	yes	6,570
305	1	yes	6,371
306	1	yes	6,268
307	0.666667	no	7,902
308	0	no	7,902
309	1	yes	4,894
310	1	yes	4,773
311	1	yes	7,048
312	1	yes	4,620
313	1	yes	6,243
314	1	yes	6,362
315	0	no	7,902
316	0	no	7,902
317	0.666667	no	7,902
318	0.333333	no	7,902
319	0.666667	no	7,902
320	1	yes	7,166

What ran

• Official BEHAVIOR v3.9.2 checkout b1979916ec1549b10a4e65e630bc6504a9af1b00, Task 01 picking_up_trash, public_test, indices 0–19, one rollout each, seed 0.
• Official horizon and robot/controller settings, RGBDFullResWrapper, no time-limit override. Horizon-limited episodes execute 7,902 steps. Video encoding off (official default); sampled observations and action/event logs retained locally.
• Frozen candidate: manifest, policy source, and final launcher.
• Batch 000 completed 301–303, then an unexpected interleaved reset caused a transport failure during 304. No completed score existed for that interrupted 304. The runner was changed to choose an unused loopback port; policy code, map, weights, seed, observations, and horizon did not change. Batch 001 resumed only missing instances 304–320. Transport record. Concurrent-client interference is suspected, not established.
• Shutdown: the official evaluator returned 0 and all 20 result JSONs were complete. The outer tool terminal session reported 143; its cause is not established. No policy/evaluator processes remained. A fresh completed-run resume check returned 0 without launching simulation.
• Simulator results were produced on the laptop RTX 5080. The deployable copy on 5080-1:/shared_work/markhsp/gpt6_b1k passed 31 CPU tests and offline model initialization; remote GPU evaluation has not been performed.

Runtime design and allowed information

The user authorized storing the scene shared by all instances. The frozen map and 57-view reference atlas were prepared from 279 stable survey frames. Offline debug poses calibrate the survey and mask task objects; task-object coordinates are not exported to runtime assets. Each episode localizes through local ALIKED/LightGlue matching, metric geometric verification, and short RGB-D tracking. Routing uses the fixed static map plus a live depth layer rebuilt on observation. This avoids carrying stale wall geometry across pose corrections. It does not yet fully solve stale object-target memory, as instance 308 demonstrates.

Runtime receives only official RGB, linear depth, proprioception, and camera-relative calibration, plus those explicit static assets. It receives no instance ID, live simulator pose, object pose, or Q. Models and dependencies are supplied locally; there are no Codex, cloud model, or LLM calls. The evaluator talks to the policy over a local loopback websocket. See PRIOR_DESIGN.md for algorithm sources, calibration provenance, and dependency licenses.

Remaining failures

• 307 (Q=0.666667): Three head-camera reacquisition failures for the final can. The recorded image visibly shows a sideways blue can. Cause of tipping not established.
• 308 (Q=0.000000): Repeated bin reacquisition failure; early stored bin position remained stale after localization improved.
• 315 (Q=0.000000): A dining-table/chair-region proposal was retained as the bin; the real bin is visible farther away near the kitchen cabinets. Two delivery-view reacquisitions failed.
• 316 (Q=0.000000): Two delivery-view bin reacquisition failures; retained target differs from the separate initial bin location.
• 317 (Q=0.666667): Only two can tracks were discovered and both were deposited. Search ended in finished_attempts without discovering the third can.
• 318 (Q=0.333333): One can deposited. Remaining target attempts exhausted after repeated head RGB-D reacquisition failures and a collision-clear grasp-stance failure. The orange can is visible in the stance-failure image; no can is visible in the first reacquisition-failure image.
• 319 (Q=0.666667): Two cans deposited. The remaining target exhausted head RGB-D reacquisition attempts at steps 6488 and 6968; no can is visible in the first failed-reacquisition image. Earlier depth-alignment and localization interruptions were recovered.

Evidence and diagnostic-only pose comparisons: failure notes. No policy change or selection of a better completed retry was made during this evaluation.

Earlier checks and limits

31 local and remote CPU checks passed. Network weight loading was disabled for offline localization/model-loading checks. Development replay of the repeated-floor-alias failure accepted 29/35 views, with maximum accepted translation error 2.85 cm; this is a diagnostic localization result, not a measured error bound for the full suite. Targeted simulator trials and earlier unsuccessful candidates are documented separately in EXPERIMENTS.md; their scores are not mixed into this table.

The deployed copy passed a final remote verification: all 673 frozen policy/asset/dependency files, the launcher, and the three official evaluator/config files match the evaluated versions. The remote archive contains all official result JSONs, source snapshot, action/event logs, and selected failure RGB-D observations; the complete sampled observation history remains on the laptop.

Run on 5080-1

cd /shared_work/markhsp/gpt6_b1k
source env_5080.sh
"$B1K_PYTHON" run.py --all --output runs/task01

For a targeted check use --indices 16 for instance 317, etc. Repeating a command with unchanged code/settings resumes missing instances; changing the policy requires a new output directory. The existing OmniGibson/Isaac Sim environment, licensed assets, and GPU are required. No Codex service or API key is required.