voxel-chain: BEHAVIOR Challenge 2026 submission

Planner + perception policy for the BEHAVIOR-1K 2026 challenge (University of Michigan). The policy follows each task's demonstrations for step order, hands and base stance, plans every motion with cuRobo against a voxel map built from the robot's own depth cameras plus a prior map of the scene's static geometry, and finds task objects by fitting their meshes to depth. No learned policy runs at test time.

This dataset holds the evaluation README (this file), the self-evaluation results and the rollout videos.

1. Policy image

Image	ghcr.io/chahyon-ku/voxel-chain-policy:final (public, no login)
Digest	sha256:454100bb503cd7d51091b2841c42ca0a21a8e5c5e0ee9701f27c9bc07612674e
Server	websocket policy server, BEHAVIOR's omnigibson.eval protocol, no auth
Port	$PORT (default 8000), listens on 0.0.0.0
Resources	1 GPU (peak ~4.6 GB VRAM measured), ~16 CPUs recommended
Environments	one environment per server (--num-envs 1); the server resets fully at each episode

docker pull ghcr.io/chahyon-ku/voxel-chain-policy:final
docker run --rm --gpus '"device=0"' --network host -e PORT=8000 ghcr.io/chahyon-ku/voxel-chain-policy:final
# the image's CMD: python scripts/07_submission/01_serve.py mode=task host=0.0.0.0 port=$PORT

Wait for the server's log line [server] listening on 0.0.0.0:8000 (after the planners warm up) before starting the evaluator.

2. Evaluator command

The stock evaluator with the stock robot config and the full-resolution RGB-D wrapper:

python -m omnigibson.eval.eval \
    --task-name <task> \
    --robot-config omnigibson/eval/r1pro.yaml \
    --env-wrapper omnigibson.eval.wrappers.RGBDFullResWrapper \
    --mode public_test --instance-indices <index> \
    --host <server host> --port 8000 --write-video

The policy reads only what that wrapper sends (proprioception, the three cameras' RGB and linear depth, camera poses and task_id). It needs no privileged state. The image runs on both BEHAVIOR-1K v3.9.2 and the v3.9.3 series.

3. Self-evaluation (public_test instances 0-9)

One rollout per instance on BEHAVIOR-1K v3.9.2 (698b27388) with the bundled r1pro.yaml, mode=task, the image's defaults. Instances without a rollout count as 0.

#	task	rollouts run	mean Q over 10 instances
0	turning_on_radio	10	1.000
1	picking_up_trash	10	0.733
2	putting_away_Halloween_decorations	0	0.000
5	setting_mousetraps	10	0.433
8	rearranging_kitchen_furniture	2	0.000
9	putting_up_Christmas_decorations_inside	0	0.000
12	preparing_lunch_box	1	0.000
16	moving_boxes_to_storage	7	0.000
18	tidying_bedroom	2	0.000
22	putting_shoes_on_rack	10	0.200
27	sorting_household_items	0	0.000
32	wash_a_baseball_cap	0	0.000
34	hanging_pictures	10	0.000
35	attach_a_camera_to_a_tripod	10	0.600
54	putting_away_toys	10	0.400
56	make_rose_centerpieces	10	0.250
59	organizing_art_supplies	10	0.120
63	installing_smoke_detectors	10	0.800
66	turning_out_all_lights_before_sleep	10	0.200
76	dispose_of_glass	10	0.425
77	installing_a_modem	10	0.425
82	store_batteries	10	0.367
83	store_honey	10	0.900
87	collecting_aluminum_cans	10	0.333
89	installing_a_fax_machine	9	0.000
90	composting_waste	10	0.050
97	setup_a_bar_for_a_cocktail_party	10	0.235

4. Files

• results/voxel_chain_public_test.zip: every rollout's eval JSON exactly as the evaluator wrote it and the per-episode table (table.md).
• videos/: each rollout's video as the evaluator wrote it (--write-video), named <task>_<instance id>_0.mp4.