troy91/behavior1k-2026-metrics

BEHAVIOR-1K 2026 Challenge: Final Submission

Submission method: Docker-based evaluation. OmniGibson (BEHAVIOR-1K v3.9.3-post2) runs outside the container and connects to our policy server with the stock websocket policy client.

path	what it is
metrics/	1,000 per-rollout result JSONs, as written by the evaluator (unmodified)
r1pro.yaml	Robot config, passed via --robot-config
wrapper.py	Env wrapper, passed via --env-wrapper

The 1,000 matching rollout videos (<task>_<instance_id>_<rollout_id>.mp4, from --write-video, unmodified) are submitted separately through the submission portal.

Metrics

metrics/ holds <task>_<instance_id>_<rollout_id>.json for 100 tasks x 10 public_test instances (301-310) x 1 rollout = 1,000 rollouts. They were produced with --mode public_test, --num-envs 1 and evaluator defaults otherwise.

metric	value
Success rate	304 / 1,000 = 30.4%
Mean final Q-score (q_score.final)	0.4808

All 100 tasks have 10 instances each, so the per-task average equals the per-rollout average.

Robot config

r1pro.yaml is an unmodified copy of the bundled omnigibson/eval/r1pro.yaml at BEHAVIOR-1K v3.9.3-post2 (byte-identical).

• Default controllers. The action is the 23-dim R1Pro action defined by this config; the policy server always returns a 23-dim action.
• Cameras (eval.camera_sensor_names): head zed_link, left/right wrist realsense_link.
• Proprioception: the config's proprio_obs list (61-dim robot_r1::proprio).
• The config's obs_modalities (proprio, rgb) and sensor_config resolution (1080) are overridden at env creation by the wrapper's camera_spec() (see below). This is the evaluator's standard behaviour.

Wrapper

wrapper.py is an unmodified copy of omnigibson/eval/wrappers/rgbd_full_res_wrapper.py at BEHAVIOR-1K v3.9.3-post2 (byte-identical), i.e. omnigibson.eval.wrappers.RGBDFullResWrapper.

• Modalities: rgb + depth_linear (plus the config's proprioception).
• Resolution: head 720x720, wrists 480x480.
• Nothing else is sent to the policy (no privileged / task::* keys).
• This is not the evaluator default (DefaultWrapper), so --env-wrapper must be passed. The same wrapper is used for all 100 tasks. The scripted-recovery (RSI) policies need the full-resolution RGB-D; the other policies resize RGB to 224x224 internally and ignore depth.

Policy server (Docker image)

• Image: ghcr.io/troylhy1991/amba-b1k-policy-server:v2 (public, no login needed; manifest digest sha256:99f1eef2ff9202dc5f02f7bcaa286233e6fb41261037cc309bb6f06cdf84aa3e).
• One container serves all 100 tasks. All checkpoints are baked in; nothing is downloaded at startup.
• Needs one GPU with 24 GB (e.g. RTX 3090 / A5000).
• Listens on 0.0.0.0:8000: websocket policy endpoint + GET /healthz (returns OK immediately, before any model is loaded).

docker pull ghcr.io/troylhy1991/amba-b1k-policy-server:v2
docker run --rm --gpus '"device=0"' -p 8000:8000 --shm-size=16g \
    ghcr.io/troylhy1991/amba-b1k-policy-server:v2

Running the evaluator against it

From the BEHAVIOR-1K OmniGibson/ directory (v3.9.3-post2), once per task:

python -m omnigibson.eval.eval \
    --task-name <task_name> \
    --robot-config omnigibson/eval/r1pro.yaml \
    --env-wrapper omnigibson.eval.wrappers.RGBDFullResWrapper \
    --mode hidden_test \
    --instance-indices <i> --num-envs 1 \
    --host <policy_container_host> --port 8000 \
    --output-dir <output_dir> \
    --write-video

• --robot-config omnigibson/eval/r1pro.yaml = r1pro.yaml; --env-wrapper omnigibson.eval.wrappers.RGBDFullResWrapper = wrapper.py.
• --num-envs 1 is required. The policy servers reject batched observations (batch > 1).
• Use evaluator defaults for everything else: --max-steps (default), --num-rollouts 1, --replay-action-chunk-size 0.
• The server picks the policy from the task_id the evaluator adds to every observation; no task argument is passed to the container.
• First observation of a new task blocks while that task's policy loads (up to a few minutes). The server keeps answering websocket pings meanwhile, so the evaluator's default client (ping timeout 300 s) does not need changes. Later episodes of the same task reuse the loaded policy; switching to another task unloads the previous one.
• One container holds one task's policy at a time. For parallel evaluation, run one container per GPU and point each evaluator process at its own container/port. Evaluating tasks in order (all instances of a task back to back) avoids reloading.