AI TrackerFrontier desk · AEST
Back to news
normalaitech

One Policy for 100 Household Tasks: Our Solution for the 2026 BEHAVIOR Challenge

arXiv cs.AI · 11 Nov 2026, 04:06 AEST

The solution for the 2026 BEHAVIOR Challenge involved a single vision-language-action model for 100 household tasks. The team achieved 9th place on the leaderboard, being the only independent team in the top 10.

Key facts

Read aloud

The solution for the twenty twenty six BEHAVIOR Challenge involved a single vision-language-action model for one hundred household tasks. The team achieved 9th place on the leaderboard, being the only independent team in the top ten. We describe our solution for the twenty twenty six BEHAVIOR Challenge: one hundred long-horizon household tasks in the OmniGibson simulator. The core of the solution is a single vision-language-action VLA model, one checkpoint for all one hundred tasks, built on our 1st-place twenty twenty five solution, which is based on $π zero point five $. Our main conclusion from twenty twenty five was that the model was undertrained, so most of the twenty twenty six work went into training the model for longer, optimizing some training parameters, making each training step slightly faster and reducing the gradient noise. A cloud evaluation infrastructure keeps a shared record with provenance, episode diagnostics and restorable simulator states, and runs six simulators on one G P U together faster than real time. Across these fourteen tasks, Astra's mean Q is zero point six three nine, against zero point one zero eight for the learned policy.

Sources in this cluster