AkashKarnatak/sentience-b1k

2026 BEHAVIOR Challenge submission

Team: Sentience | Affiliation: Sentience Inc | Date: 2026-10-16

Policy

The policy is modified pi0.5 fine-tuned on the 2026 challenge demonstrations with an event-anchored progress head. The policy predicts its own sub-goal count and receives it back as one state token, so long multi-object tasks are tracked without a separate planner. Several fine-tuned checkpoints of the same model are served from one image and the server selects the checkpoint for the evaluator's task from the observation's task_id.

Policy inputs: head RGB (720x720), left/right wrist RGB (480x480), proprioception (61-d), task id. Depth is received and not used. No privileged information of any kind.

Setup

Requirements: one NVIDIA GPU with 24 GB of memory, an NVIDIA driver that supports CUDA 12 (CUDA and cuDNN are inside the image), about 90 GB of free disk for the image, and port 8000 free. The simulator runs outside the container.

Step 1: pull the image

docker pull docker.io/akashkarnatak/sentience-b1k:v5

Image digest: sha256:3f660d20bab743c9d4b4c37b737540c50eeff341826c02417db25bccc28bb263

Step 2: start the policy server

docker run -d --name sentience-b1k-policy --gpus '"device=0"' -p 8000:8000 docker.io/akashkarnatak/sentience-b1k:v5

The server listens on port 8000 (WebSocket policy protocol of the BEHAVIOR evaluator, msgpack messages).

Step 3: wait until the policy is up

# This may take around 5 to 6 minutes: the server loads and compiles the multiple checkpoints the policy
# serves before it reports ready. Please be patient; the evaluator can be started once this loop finishes.
until curl -sf http://127.0.0.1:8000/healthz > /dev/null; do sleep 5; done; echo "policy is up"

Progress can be followed with docker logs -f sentience-b1k-policy.

Step 4: evaluate

With BEHAVIOR-1K v3.9.3-post1 installed on the host, run the official evaluator against the container for each task:

python -m omnigibson.eval.eval \
  --task-name <task> --instance-indices 0 1 2 3 4 5 6 7 8 9 --num-rollouts 1 \
  --host 127.0.0.1 --port 8000 \
  --env-wrapper omnigibson.eval.wrappers.RGBDFullResWrapper \
  --output-dir <out>/<task> --write-video

One container serves all 100 tasks, in any order. Run one evaluator process per container at a time; for parallel evaluation start one container per GPU on different host ports.

To stop the server: docker rm -f sentience-b1k-policy.

Wrapper and robot config

• Wrapper: omnigibson.eval.wrappers.RGBDFullResWrapper, bundled with BEHAVIOR-1K and unmodified (copy: wrapper/rgbd_full_res_wrapper.py).
• Robot config: OmniGibson/omnigibson/eval/r1pro.yaml, bundled and unmodified (copy: robot/r1pro.yaml).
• Action space: the default R1Pro 23-d action (base velocity, torso, arms, grippers); --max-steps not set (default 1.5x mean human demo length); single environment per evaluator process; tiled rendering off; headless.

Results

• json/: 1000 metrics files (<task>_<instance_id>_0.json, instance ids 301-310 = public-test indices 0-9), written by the evaluator above and not modified. 100 tasks x 10 instances x 1 rollout, OmniGibson v3.9.3-post1, evaluated 2026-10-16 with the policy served from this Docker image. Download (submission package, zip): https://huggingface.co/datasets/AkashKarnatak/sentience-b1k/resolve/main/standard.public.Sentience.Sentience_Inc.20261016.zip?download=true
• Videos: one MP4 per rollout (<task>_<instance_id>_0.mp4, written by the evaluator with --write-video, unmodified), uploaded separately as sentience-submission-video-rollouts-b1k.zip. Download (videos, zip): https://huggingface.co/datasets/AkashKarnatak/sentience-b1k/resolve/main/sentience-submission-video-rollouts-b1k.zip?download=true
• Overall q_score 0.3788, task success rate 0.1860 (186/1000 successes).