One VLA policy places 9th in the 2026 BEHAVIOR Challenge
Original title: One Policy for 100 Household Tasks: Our Solution for the 2026 BEHAVIOR Challenge
A single π0.5-based checkpoint scores 0.392 Q on 100 tasks; GPT-6 Astra averages 0.639 Q on 14 routed tasks.
An independent team describes a 9th-place entry, the only independent team in the top 10, for the 2026 BEHAVIOR Challenge of 100 long-horizon household tasks in OmniGibson. One vision-language-action checkpoint, built on their 2025 π0.5-based winner, covers all 100 tasks. Shared-task Q rises from 0.260 to 0.371, and all-task Q is 0.392 after longer training on about 211 million demonstration frames. Fourteen tasks are routed to a GPT-6 Astra coding agent; the public submission included only part of those results and scored 0.446 Q with 27.3% full-task success.
Unconfirmed: Leaderboard placement and scores are as stated by the team.
- On 50 tasks shared with 2025, Q rises from 0.260 to 0.371.
- Training is a 200k-step run, about one pass over 211 million frames.
- Astra's mean Q on 14 tasks is 0.639 versus 0.108 for the learned policy.
- The public score is 0.446 Q with 27.3% full-task success.
- Six simulators run on one GPU faster than real time.
- Products
- OmniGibson
- AI models
- π0.5GPT-6 Astra
From the original article
1 min read at arxiv.orgWe describe our solution for the 2026 BEHAVIOR Challenge: 100 long-horizon household tasks in the OmniGibson simulator. We placed 9th on the final leaderboard, the only independent team in the top 10. The core of the solution is a single vision-language-action (VLA) model, one checkpoint for all 100 tasks, built on our 1st-place 2025 solution, which is based on $π_{0.
By Ilia Larchenko · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.