One learned policy, plus an agent on selected robot tasks

Household activities can ask a robot to move through a room, grasp objects, operate appliances and keep track of a task unfolding over many steps. The 2026 BEHAVIOR Challenge tested these demands in OmniGibson, a simulator, using a Galaxea R1Pro robot and 100 activities, from turning on a radio to cooking a meal. Full success required satisfying every goal condition. Otherwise, the Q score measured the fraction of conditions satisfied.

In an arXiv preprint, Ilia Larchenko and colleagues report placing ninth overall and first among independent teams. Their system used one vision-language-action (VLA) model—a model that combines visual and robot-state information to produce actions—and a GPT-6 Astra coding agent on selected tasks. The learned model, built on the team’s 2025 solution based on π0.5, used one checkpoint for all 100 tasks; the submitted system routed 14 tasks to Astra.

Adapting the learned policy

The team’s 2025 entry used four checkpoints across 50 tasks. For 2026, the researchers adapted the model to cover 100 tasks and trained it for 200,000 steps, roughly one pass over the demonstrations. The learned policy scored 0.392 across all 100 tasks. On the 50 tasks shared with the previous challenge, its Q score rose from 0.260 to 0.371.

The team changed how the model represented task stages, using skill annotations from demonstrations instead of dividing episodes into equal time segments. New task and stage representations were initialized using similar tasks from the earlier system. The researchers also narrowed the action space to 21 dimensions matching the controls used by human teleoperators and used a sampling approach intended to reduce training noise.

These changes formed a combined recipe, not a controlled test of each component. The authors say they cannot attribute the improvement on shared tasks to longer training alone: the model also incorporated demonstrations from 50 additional tasks, model changes and bug fixes. Evaluations used different simulator versions, though the team reports seeing no meaningful performance difference between them.

An agent on selected tasks

Astra controlled the robot through observation and motion tools. It inspected the scene and issued movement goals rather than sending low-level motor commands directly. Before evaluation, it practiced on training instances, where it could reset scenes, inspect goal progress and consult demonstrations. The team then froze each practice session and its notes. Evaluation episodes began from a fresh copy, without the privileged practice tools.

The researchers routed 14 tasks to Astra after trials suggested it might perform well on them. Across those selected tasks, Astra’s mean Q was 0.639, compared with 0.108 for the learned policy. This is not a comparison on a shared assignment: the agent’s tasks were selected after favorable trials, and the scores do not show how the two systems would compare across all 100 tasks under the same conditions. Practice, routing and tuning also took place close to the deadline.

The public submission scored 0.446 Q and had a 27.3% full-task success rate, but included only some of Astra’s results. The authors say the organizers did not re-evaluate their entry on the private test set; the reported score came from the team’s own evaluation using the official evaluator. Astra was also slow and costly: an episode took a median of 0.63 hours and cost tens of dollars in API fees. The learned policy alone scored 0.392 across all 100 tasks.

Performance varied by task. The learned policy scored at least 0.75 on 13 tasks and zero on eight. In the team’s failure review, dexterity remained a common weakness. The authors also describe the policy recovering from some mistakes and adapting familiar actions, but these observations do not establish how reliably that behavior would transfer to new settings.

The findings are limited to a competition evaluation in a simulator, with a particular robot, dataset and scoring system. The paper is an arXiv preprint; the provided source does not establish whether it has been peer reviewed. The reported results describe this team’s system and do not show that it is ready for everyday use or that the same approach would work in real homes.