One Policy for 100 Household Tasks: Our Solution for the 2026 BEHAVIOR Challenge
AuthorsIlia Larchenko, Gleb Zarin, Atindra Nair, Keerti Manney, Vitaly Kleban, Mirko Kiefer, Anjana Yodaiken
A single vision-language-action policy tackles 100 simulated household tasks, with a coding agent stepping in for some of the hardest ones.
Key results
One learned-policy checkpoint covers the full BEHAVIOR task set.
Length of the final policy training run.
Score across all 100 tasks.
Improved from 0.260 in the 2025 submission.
Across 14 selected tasks, versus 0.108 for the learned policy.
Success rate for the public submission, which scored 0.446 Q.
What the paper found
This solution to the 2026 BEHAVIOR Challenge uses one π0.5-based vision-language-action policy for all 100 household tasks in OmniGibson, which runs on NVIDIA Isaac Sim. To expand from the previous 50-task setup, it replaces time-based stages with skill-annotation-driven semantic stages, warm-starts new task and stage embeddings from similar tasks, and maps robot actions to a 21-dimensional control space. Training improvements include stratified flow-matching times and antithetic noise to reduce gradient variance, followed by a 200K-step run. The learned policy scores 0.392 across all tasks; on the 50 tasks shared with 2025, its Q rises to 0.371 from 0.260. For 14 tasks where the policy was weak, the team instead routed control to OpenAI’s GPT-6 Astra, which practiced on training instances and then acted through observation and motion tools. Astra averages 0.639 Q on those tasks, compared with 0.108 for the learned policy. The public submission placed ninth, scoring 0.446 Q with 27.3% full-task success. The main trade-off is that the language-model agent improves performance on selected tasks but runs slowly and incurs substantial API costs.
Original 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.
Read the original paperMore in Embodied AI
Browse all 69 papers →Demonstration of MANA: Multi-perspective Alignment of NPC Agents
Amir Tahmasbi, Keshav Singh, Shaheer Raza, Owen Mei, Jason Wu
MANA lets game developers step into an NPC’s shoes and demonstrate how it should behave, then lets the character act on its own.
Kimodo-Policy: Learning Kinematic Action Models for Humanoid Loco-Manipulation
Yunheng Wang, Jiaxi Zhang, Wei Lu, Yuetong Fang, Taowen Wang, Yixiao Feng, Shuning Zhang, Junzhe Xu, Yichi Wang, Zizhao Yuan, Kailun Huang, Lusong Li, Mingqian Zhou, Zecui Zeng, Renjing Xu
Kimodo-Policy turns a pretrained whole-body motion generator into a control policy for humanoid robots that walk and manipulate objects.
JustRUA: A Controlled Study of GPT-6 Robot-Use Agents
Zefan Wang, Yufan Dang, Xuanwei Shi, Mingqi Gao, Cheng Yin, SikYuen Tam, Hanyu Liu, Yasheng Wang, Yaxi Lu, Songyang Zhang, Xin Cong, Yuan Yao
A leaner harness helps a GPT-driven robot perform much better in simulation, showing that agent design choices can matter as much as learning.