SakamotoRyoma/1014.5

BEHAVIOR-1K 2026 evaluation instructions

Policy server submission. No Docker image.

Policy server

• URL: ws://58.250.250.24:19500
• Protocol: plain WebSocket (ws://), no TLS and no authentication
• Serialization: the official OmniGibson WebSocket/MessagePack protocol
• Ports: 19500 through 19599 inclusive (100 TCP ports)
• Use one port per parallel rollout. The URL above is only the first port. Assign each parallel rollout a distinct port in this range. A new connection to a port replaces any existing connection on that port.
• Action-chunk replay: K=10, enabled with --replay-action-chunk-size 10. Every task supports K=10.
• Observations: the server accepts both batched observations (leading batch dimension of size 1) and unbatched observations.

Capacity

• A rollout that cannot get capacity within 60 seconds of its first observation fails. Keep concurrency within these limits.
• Reuse a port for consecutive rollouts: run the next instance on the same port. A port that is left idle without a reset keeps its capacity reserved for up to one hour.

Evaluator command for the joint test

Use the official BEHAVIOR-1K v3.9.3-post1 evaluator.

python -m omnigibson.eval.eval \
  --task-name <task_name> \
  --host 58.250.250.24 --port <port> \
  --mode public_test --instance-indices <instance_index> \
  --num-rollouts 1 --num-envs 1 \
  --replay-action-chunk-size 10 \
  --output-dir <output> --write-video \
  --env-wrapper omnigibson.eval.wrappers.DefaultWrapper \
  --robot-config robot/r1pro.yaml \
  --headless

Replace all placeholders before running:

• <task_name>: the BEHAVIOR task name.
• <port>: one integer from 19500 through 19599.
• <instance_index>: one integer from 0 through 9.
• <output>: the output directory for metrics and videos.

With --num-envs 1, evaluate one instance at a time. To evaluate all ten instances, run the command separately for each index, on the same port.

The stock v3.9.3-post1 evaluator sends observations with a leading batch dimension of size 1 even with --num-envs 1. The organizer-side test is planned to use unbatched observations, as previously discussed. The server accepts either format.

Official rules: https://behavior.stanford.edu/challenge/submission.html

Connectivity and reconnection test

Please coordinate with us before sending observations or reset messages, and before triggering a disconnect. We will designate a dedicated port for the reconnection test.

A health check is available at:

http://58.250.250.24:<port>/healthz

For the reconnection test, use an organizer-side client with reconnection enabled; the stock evaluator command above does not enable automatic reconnection.

The test should verify that, after a disconnect following inference but before the response is sent:

• The client reconnects to the same port and resends the outstanding observation.
• The rollout resumes without a reset.
• Server-side history is retained, and the repeated observation does not update history twice.
• A reset is sent before a new rollout.

The resent observation must be byte-identical to the outstanding one and must be the first message on the new connection.

Evaluation materials

• metrics/: unmodified official metrics JSON, one file per rollout (1000 files).
• videos/: unmodified official MP4 files, one per rollout (1000 files). Videos are in the dataset videos/ directory, not inside the zip.
• robot/r1pro.yaml: the robot config passed as --robot-config (sha256 prefix a98fa8a81472adfe).
• wrapper/default_wrapper.py: the official omnigibson.eval.wrappers.DefaultWrapper (224x224 RGB).

Results archive: submission_metrics_readme_robot_wrapper.zip