Predicate-Aware VLA for the 2026 BEHAVIOR Challenge
An auxiliary progress head on the π₀.₅ vision-language-action policy for the 2026 Stanford BEHAVIOR Challenge, tested in a pre-registered A/B on rented GPUs. The head learns without degrading the action loss; the A/B found no significant difference between arms.
An entry for the 2026 Stanford BEHAVIOR Challenge, built with one teammate on a $200 budget of rented GPUs. The contribution is an auxiliary progress head on π₀.₅: a second output that predicts how much of the task’s goal is already satisfied, trained alongside the action decoder so the shared representation has to encode task progress rather than only the next action.
What was measured:
- Baseline: the released checkpoint scores Q = 0.111 on
turning_on_radio(3 of 27 instances). - Training: both arms trained for 10,000 steps; the progress head learned (loss 0.694 → 0.563) without degrading the action loss.
- Pre-registered A/B: no significant difference. On the radio task, arm A 38/108 against arm B 40/108 (ΔQ +0.019, 95% CI −0.091 to +0.128, p = 0.73).
- Test set: a policy we trained completes the radio task on 9 of the 20 public test instances. The leaderboard averages over 100 tasks, so the overall score is about 0.0045.
The repository also contains the parallel, resumable rollout harness, the dataset slicing and progress-label pipeline, and the logs behind every result.