Back to AI News
arxiv

Single VLA checkpoint tackles 100 household tasks in BEHAVIOR 2026

The team describes its 2026 BEHAVIOR Challenge solution: one vision-language-action checkpoint for all 100 long-horizon household tasks in OmniGibson, built on its 2025 entry based on pi_0.5. It placed 9th, reporting Q 0.371 on the 50 tasks shared with 2025 versus 0.260 previously, and 0.392 across all 100 tasks. The work focused on longer training, a 21-dimensional action space, variance-reduced multi-sample flow matching, and cloud evaluation infrastructure. For 14 tasks, the system routes control to a GPT-6 Astra coding agent, which achieved mean Q 0.639 versus 0.108 for the learned policy.

AI Brief

It provides a concrete technical account of scaling one VLA policy across many long-horizon robot tasks and combining it with an agentic coding model.

arXiv:2611.14432v1 Announce Type: cross Abstract: We 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 $\pi_{0.5}$. On the 50 tasks shared with 2025, its Q rises from 0.260 to 0.371; across all 100 tasks it scores 0.392. Our main conclusion from 2025 was that the model was undertrained, so most of the 2026 work went into training the model for longer, optimizing some training parameters, making each training step slightly faster and reducing the gradient noise: a migration from 50 to 100 tasks (semantic stages built from the demonstrations' skill annotations, task and stage embeddings of new tasks warm-started from similar old ones, and a 21-dimensional action space matching the teleoperation controls); variance-reduced multi-sample flow matching with a simple noise-vs-compute model that picks the number of action-expert samples per observation without ablations; and a 200k-step training run, about one pass over the 211M frames of the demonstrations. A cloud evaluation infrastructure keeps a shared record with provenance, episode diagnostics and restorable simulator states, and runs six simulators on one GPU together faster than real time. On top of the learned policy, 14 tasks are routed to a GPT-6 Astra coding agent that controls the robot through observation and motion tools, starting every episode from a frozen copy of its own practice session on training instances. Across these 14 tasks, Astra's mean Q is 0.639, against 0.108 for the learned policy. The public submission included only part of the Astra results and scored 0.446 Q with 27.3% full-task success.
Be the first to rate this article.

Comments & Questions (0)

Captcha
Please be respectful — let's keep the conversation friendly.

No comments yet

Be the first to comment!