BEHAVIOR 2026 VLA
One vision-language-action (VLA) checkpoint for all 100 household tasks of the 2026 BEHAVIOR Challenge, running a Galaxea R1Pro robot in the OmniGibson simulator.
The model is based on Pi0.5 and warm-started from our 2025 solution.
- Code: behavior-1k-solution-2026
- Technical report: TODO
Results
Local evaluation with the official success criterion and q-score: 100 tasks, 10 episodes per task (instances 301-310), learned policy only.
| Tasks | Count | q-score |
|---|---|---|
| All | 100 | 0.389 |
| Shared with the 2025 challenge | 50 | 0.370 |
| New in 2026 | 50 | 0.407 |
The success rate over all tasks is 21.8%. The q-score counts a full success as 1 and otherwise gives partial credit for the fraction of satisfied goal predicates.
Files
| Path | Contents |
|---|---|
params/ |
EMA model weights in Orbax format, 10.7 GB |
assets/b1k_2026/norm_stats.json |
State and action normalization statistics, action correlation matrices |
assets/b1k_2026/fast_tokenizer/ |
FAST action tokenizer, used for the auxiliary training loss |
Usage
Model
- Inputs: three 224x224 RGB cameras (head and both wrists), robot state, task, current stage
- Outputs: 30-step action chunks over 21 degrees of freedom: base (3), trunk (2), two arms (7 each), two grippers (1 each)
- Architecture: PaliGemma vision-language model and a smaller action expert trained with flow matching, 2.9B parameters
- Conditioning: learned embeddings for 100 tasks and 413 semantic stages (2-9 per task), no text prompts
Training
Trained for 200k steps on the 20,000 demonstrations of the 2026 challenge dataset, warm-started from a converted 2025 checkpoint. The optimizer was AdamW with EMA 0.995, using 10 variance-reduced flow samples per observation and auxiliary stage and FAST losses.
This policy is only one part of our solution. The second part, covering the 14 tasks where the policy scores lowest, is described in the code repository and the technical report.