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

Ilia Larchenko, Gleb Zarin, Atindra Nair, Keerti Manney, Vitaly Kleban, Mirko Kiefer, Anjana Yodaiken

November 2026

Head-camera view of a simulated rollout at a kitchen counter with a cutting board.

TL;DR One pi0.5-based VLA checkpoint, trained for 200k steps on 20,000 demonstrations, covers all 100 BEHAVIOR household tasks in simulation and scores 0.392 Q; on the 50 tasks shared with 2025, Q rises from 0.260 to 0.371. Routing 14 hard tasks to a GPT-6 Astra coding agent that practiced beforehand lifts the public submission to 0.446 Q.

What they do

A report on the 9th-place entry (the only independent team in the top 10) to the 2026 BEHAVIOR Challenge: 100 long-horizon household tasks for a Galaxea R1Pro in the OmniGibson simulator, with 200 teleoperated demonstrations per task. A single vision-language-action checkpoint handles all 100 tasks, and 14 tasks where it is weak go to an LLM coding agent instead.

How it works

Growing 50 tasks to 100. The team’s 2025 pi0.5-based model is converted, not retrained: semantic stages built from skill annotations (413 instead of 1,093 time slices), new task and stage embeddings initialized from similar old tasks, and a 21-dimensional action space matching the teleoperation controls.
Less gradient noise. Each observation gets K action-expert flow samples with stratified flow times and antithetic noise; a noise-versus-compute model picks K of about 10 from one gradient-variance probe.
Training and evaluation. 200k steps, about one pass over 211M frames. Evaluation runs six simulators per GPU at 1.11x real time in aggregate and saves restorable simulator states.
Agent route. GPT-6 Astra practices each task on training instances, then starts every evaluation episode from a frozen copy of that session and controls the robot through observation, depth-position and goal-motion tools.

Results

The learned policy alone scores 0.392 Q on 100 tasks. On the 50 tasks shared with 2025, Q rises from 0.260 to 0.371 despite a 25% shorter time limit.
On the 14 agent tasks, Astra reaches a mean Q of 0.639 against 0.108 for the policy. The public submission scored 0.446 Q with 27.3% full-task success.
On ten tasks reviewed in both years, full success rose from 20% to 36%; dexterity remained the most common failure reason.

Limitations

Simulation only, and the score is self-reported on public instances. The registers, sinks and scratch tokens were not ablated, and the gain over 2025 cannot be attributed to longer training alone. Eight tasks score zero. An agent episode takes a median of 0.63 h of wall time, about 17 times longer than real time, and costs tens of dollars in API fees; only 106 of 140 agent episodes made the public results.

Videos

Wrist and head camera views of a simulated rollout at an open fridge.

Head-camera view of a simulated rollout at the kitchen oven and fridge.

Figures

From Figure 1: The 2026 model. The VLM (left) encodes three camera images, register and scratch tokens, the task and ... — From Figure 1: The 2026 model. The VLM (left) encodes three camera images, register and scratch tokens, the task and stage embeddings and the robot state; the smaller action expert (right) denoises a 30-step action chunk and reads the VLM through a learnable mixture of the keys and values of all VLM layers.

From Figure 10: The Astra policy. Practice (left) uses training instances, resets, saved states and goal-predicate fe... — From Figure 10: The Astra policy. Practice (left) uses training instances, resets, saved states and goal-predicate feedback to produce a frozen session, notes and recipes. Each agent-routed episode (right) forks that session. Routing (bottom) selects Astra for 14 enabled tasks and the learned policy otherwise.

From Figure 11: Public per-task Q, ordered by the learned policy's score. On the 14 Astra tasks, stacked bars reach A... — From Figure 11: Public per-task Q, ordered by the learned policy's score. On the 14 Astra tasks, stacked bars reach Astra's Q; other bars show the learned policy's Q.

From Figure 13: Manually annotated failure reasons on ten tasks shared by the 2025 submission and the 2026 learned po... — From Figure 13: Manually annotated failure reasons on ten tasks shared by the 2025 submission and the 2026 learned policy.

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 $π_{0.5}$. 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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong