gpt6_b1k — standalone Task 01 policy
Target: mean official final Q above 0.7 on all 20 public picking_up_trash instances (301–320). Complete frozen evaluation: mean Q = 0.766667, 13/20 successes; target achieved. See EVALUATION.md for all 20 official results.
The policy runs locally using NumPy, SciPy, OpenCV, static R1Pro kinematics, RGB, linear depth, and proprioception. The fixed-scene prior mode additionally uses local Torch/ALIKED/LightGlue models and precomputed map assets. The visual navigator also uses camera-relative calibration supplied in official evaluation observations. It makes no Codex/backend/LLM calls and needs no API key. The runner selects a free loopback port for each launch; its localhost websocket connects the official evaluator to the local policy. Use --port only when a fixed port is needed. The policy never imports OmniGibson or reads scene-state files; the separate official evaluator runs the simulator. Privileged object/world geometry is confined to the separate debug/ tools.
Measured status
The default prior policy completed all 20 official instances, with mean final Q 0.766667 and 13/20 official successes. Results include every failed episode, using one frozen policy, seed 0, the official horizon, and full RGB-D observations. The complete table, infrastructure-resume provenance, and remaining failures are in EVALUATION.md.
The policy uses a user-authorized static scene map, 57 reference views, local ALIKED/LightGlue localization, short RGB-D tracking, live depth obstacle checks, and reachability-aware grasp/placement stances. The map is prepared once. Dynamic cans and bin are detected from current observations; their positions are not stored in the prior. See PRIOR_DESIGN.md.
Earlier targeted tests and failed candidates are recorded separately in EXPERIMENTS.md. Their scores are not combined with the completed frozen evaluation. This result demonstrates the agreed known-scene mean-Q target, not success on every instance or unseen-scene generalization.
Existing laptop environment
cd /home/eai/yixin/gpt6_b1k
source env.sh
Defaults use the existing behavior-392dev environment, the clean BEHAVIOR v3.9.2 checkout at /home/eai/code/BEHAVIOR_1K_v392, and assets at /home/eai/code/b1k/datasets. For another computer, set B1K_PYTHON, B1K_REPO, and OMNIGIBSON_DATA_PATH before sourcing env.sh. OmniGibson, Isaac Sim, licensed assets, and the GPU driver must already be installed. No installation or simulator modification is performed by this package.
Copy on 5080-1
ssh 5080-1 reaches TRT-EAI-OLDLAB-1. The project destination is /shared_work/markhsp/gpt6_b1k. There, use source env_5080.sh to select the existing behavior2026 Python environment, official v3.9.2 export, and shared datasets. Simulator development results reported here were produced on the laptop; the remote copy does not imply remote GPU validation. Follow the site's existing GPU allocation procedure before launching simulation there.
Short development trial
"$B1K_PYTHON" run.py --navigation prior --indices 0 --max-steps 1200 --output runs/pilot01
Explicit time-limited runs are labeled diagnostic. Each run stores a source fingerprint, commands, action logs, RGB-D/proprio snapshots, and official evaluator outputs. Use a new output directory after changing code. Repeating a command with identical code/settings skips completed instances.
prior is the CLI default. The older visual and validated modes remain available for explicit comparison. Every mode runs without an external model service.
Full evaluation command (expensive)
"$B1K_PYTHON" run.py --navigation prior --all --output runs/frozen20
This runs 20 instances sequentially in one official evaluator process, reusing the loaded simulator. It uses the official default horizon, default robot/controller settings, seed 0, and full RGB-D wrapper. Results appear under runs/frozen20/batch_000/evaluation/; policy recordings are under batch_000/policy/. A source snapshot is stored with each run. Progress summaries update after each completed instance. Full video encoding is optional with --write-video; sampled RGB-D observations and action logs are always retained. This follows the official evaluator's default video setting and does not change its observations or metrics. The target is only declared achieved if all 20 completed with mean final Q > 0.7 and no diagnostic horizon override. Resuming runs only the missing indices in a new batch and retains failed-task scores rather than selecting better retries.
Inspect without restarting simulation for every experiment
"$B1K_PYTHON" debug/session.py --instance 301 --output debug/lab301
"$B1K_PYTHON" debug/analyze.py debug/lab301/initial
The development lab accepts JSON commands in debug/lab301/command.json, e.g. {"op":"run","name":"scan01","scan_only":true,"steps":1200,"stop_phase":"inspection_complete"}. It reuses the loaded scene, resets for each experiment, and writes sensor recordings and separate privileged diagnostic geometry. End with {"op":"quit"}. These commands and geometry are never used by run.py or the final policy server.
Checks
"$B1K_PYTHON" -m pytest -q tests
Motion/sensor helpers were adapted from the user-provided 5080-1:/shared_work/markhsp/gpt6_behavior_solution reference, via its preserved r21 source. Robot geometry and a robot-only spare-arm posture are reused; no task2 scene coordinates or action replay are used.