KWSMooBang/submission_eval_mix1

BEHAVIOR-1K 2026 Challenge — submission README

Policy: G0.5 (GalaxeaVLA) with subtask chain-of-thought + skill-routed action experts, evaluated on all 100 tasks x public test instances 0 … 9 (instance files 301–310) x 1 rollout = 1,000 rollouts, default evaluator timeout (1.5x mean human demonstration length). Overall mean q_score: see summary.txt.

Contents

Path	What
json/	1,000 result files written by OmniGibson/omnigibson/eval/eval.py, unmodified (<task>_<instance_id>_0.json)
eval_wrapper/rgbd_full_res_wrapper.py	evaluation wrapper used (official omnigibson.eval.wrappers.RGBDFullResWrapper, unchanged)
robot_config/r1pro.yaml	robot config used: the bundled default omnigibson/eval/r1pro.yaml, unchanged (no --robot-config)
serving/	policy serving configuration, task prompt config, model/expert SHA256, software versions
summary.txt	per-task and overall q_score / success
SHA256SUMS	checksums of every file in this zip

Rollout videos (1,000 MP4, one per JSON, same base names) are provided through the link submitted on the portal.

Evaluator command (per rollout)

The policy server and the evaluator were started fresh for every (task, instance) rollout:

python -m omnigibson.eval.eval \
  --task-name <task> --host <policy_host> --port <port> \
  --instance-indices <i> --num-rollouts 1 \
  --env-wrapper omnigibson.eval.wrappers.RGBDFullResWrapper \
  --output-dir <out> --write-video --headless

• <i> in 0 … 9; no --max-steps (default timeout); no --robot-config (default r1pro.yaml).
• Wrapper path: OmniGibson/omnigibson/eval/wrappers/rgbd_full_res_wrapper.py (RGB + depth, head 720x720, wrists 480x480, plus proprioception). Only RGB, depth and proprioception reach the policy.
• Simulator: stanfordvl/behavior:3.9.2 image with the OmniGibson source at 3.9.3 (serving/versions.txt), Isaac Sim 5.1.0. Note: Isaac Sim 5.1 crashes at startup on NVIDIA driver 595.x; use <= 590 (we used 590.48.01 and 580.159.03).

Policy server

WebSocket policy server (GalaxeaVLA/scripts/serve_g05_b1k.py), action space = 23-D R1Pro action of the default robot config. Configuration (serving/eval_mix1_step180k.env):

• base checkpoint step_180000_model.pt + dataset_stats.json; 7 skill-group LoRA action experts (all, grasp, inplace, navigation, place, push_contact, tool), routed by the generated subtask
• task prompt config behavior_final_cot.yaml; 32-step action chunks executed before re-planning
• flow-matching action head; fp32 weights with bf16 autocast; ~17.5 GB GPU memory (fits a single 24 GB GPU)

Final-evaluation access (Docker image or IP address with >= 50 ports): TO BE FILLED IN before submitting.

Notes on the reported rollouts

• Each (task, instance) is reported with exactly one rollout; no outcome-based selection.
• Rollouts whose evaluator crashed or was interrupted before writing a result were re-run with the same configuration: 24 interrupted by a host reboot, 11 GPU out-of-memory (several simulators per GPU), 1 simulator hang at startup.
• 2 rollouts (collecting_aluminum_cans instances 301 and 309) ended without a result after a PhysX error (Illegal BroadPhaseUpdateData) 36–48 min into the episode, and were re-run; the re-runs are reported.
• 20 rollouts (laying_tile_floors, sorting_books_on_shelf, all 10 instances each) were first run with a wrong serving setting (16-step chunks instead of 32). Both tasks were re-run in full with the configuration above, regardless of the earlier outcomes; only the re-runs are reported.