naami3/behavior26-pi05-radio

behavior26: π₀.₅ with a progress head, turning_on_radio (BEHAVIOR Challenge 2026)

Weights of the policy team behavior26 submitted to the 2026 BEHAVIOR-1K Challenge, standard track.

• Code, configs, training logs: https://github.com/wimaan3/behavior26
• Technical report: report/behavior26_report.pdf in the GitHub repository
• Docker image (same weights, ready to serve): wimaan3/behavior26-policy:armB-5999@sha256:84f7bd76d2edc28eeb88c4a16dd119d7c6fd3d5de55d77b85a46bd3ba904a73e
• Authors: Imaan Soltanalipour, Jeffrey Le (no affiliation)

What this is

An openpi π₀.₅ checkpoint (orbax format, params/ only, no optimizer state) for the single task turning_on_radio on the Galaxea R1Pro robot.

• Initialised from the organisers' released turning_on_radio π₀.₅ checkpoint (pi05_b1k).
• Fine-tuned 6,000 steps, batch 32, seed 42, on the 200 turning_on_radio demonstrations of the challenge dataset, with the vision-language backbone frozen (config pi05_b1k_frozen_vlm_progress).
• Adds an auxiliary progress head: a linear layer on the mean-pooled action-expert tokens, trained with binary cross-entropy against a goal-progress label (loss weight λ = 0.15). The head is not used at inference; actions come from the unchanged flow-matching decoder.
• Norm statistics: assets/turning_on_radio/norm_stats.json, the released checkpoint's own.
• Prompt: the task name, turning_on_radio.

Results

Evaluation	Successes
Public test instances 301–320, 1 rollout each (submitted entry)	9 / 20 (45%)
Training instances 10–36, 4 rollouts each	40 / 108 (37.0%)
Same fine-tune without the progress head (control), instances 10–36	38 / 108 (35.2%)

The paired A/B difference is +0.019 (95% CI [−0.091, +0.128], p = 0.73), so the progress head made no measurable difference on this task. The control also scored 9/20 on 301–320. Only one of the challenge's 100 tasks was attempted, so the overall challenge score is about 0.0045.

How to run it

Requires the patched openpi from the GitHub repository (the progress head adds parameters that stock openpi does not know about).

git clone https://github.com/wimaan3/behavior26 && cd behavior26
git clone https://github.com/wensi-ai/openpi.git ../openpi
git -C ../openpi checkout 0cc8e355f7bac0976db1cc3139b1ff0379feea60
OPENPI_ROOT=../openpi scripts/apply_openpi_patches.sh
hf download naami3/behavior26-pi05-radio --local-dir ../ckpt
cd ../openpi && uv sync && uv run scripts/b1k/serve_b1k.py \
  --robot b1k/R1Pro --task b1k/turning_on_radio --repo-id turning_on_radio --port 8000 \
  --policy.config pi05_b1k_frozen_vlm_progress --policy.dir ../ckpt

Or skip the setup and run the Docker image above: docker run --gpus all -p 8000:8000 <image>. Then run the BEHAVIOR-1K evaluator against port 8000; the exact command is in the GitHub README. One GPU with bfloat16 support (Ampere or newer); about 6 GB of GPU memory.

Licence

π₀.₅ builds on PaliGemma, so these weights are a Model Derivative of Gemma and are distributed under the Gemma Terms of Use, included here as LICENSE_GEMMA.txt. Use is subject to the Gemma Prohibited Use Policy. See NOTICE. Our code in the GitHub repository is MIT-licensed; openpi is Apache-2.0.