One VLA Policy Tackles 100 Household Tasks: A Technical Breakdown of the BEHAVIOR Challenge

An independent team used a single VLA model plus a hybrid coding agent to place ninth among 100 tasks in the BEHAVIOR Challenge.
This article presents an independent team's solution in the 2026 BEHAVIOR Challenge. Using a single VLA model for 100 long-horizon household tasks, the team placed ninth and was the only independent entrant in the top ten. Key improvements include extending training to 200,000 steps, optimizing flow matching sampling and gradient noise, and building cloud-based evaluation infrastructure with six parallel simulators on one GPU. A hybrid architecture routed 14 tasks to a GPT-6 Astra coding agent, achieving 0.639 average Q versus 0.108 for the learned policy. The final public score was 0.446 Q with a 27.3% full-task success rate.
In the 2026 BEHAVIOR Challenge, an independent team used a single vision-language-action (VLA) model to tackle 100 long-horizon household tasks, ultimately placing ninth on the leaderboard and becoming the only independent entrant in the top ten. The core of this achievement lies in a deeply optimized policy network and a bold hybrid architecture experiment—routing some tasks to a coding agent for control.

From 50 to 100: The Engineering Challenge of Scaling Up
The BEHAVIOR Challenge requires robots to complete 100 household tasks in the OmniGibson simulator, each involving long-horizon planning and fine manipulation. The team's core strategy was to train a unified VLA model, with all tasks sharing the same checkpoint. This design continues from the 2025 winning solution, which was built on the π₀.₅ model.
On the 50 tasks shared between 2025 and 2026, the model's Q value improved from 0.260 to 0.371; across all 100 tasks, the model scored 0.392. The team's main takeaway from the 2025 experience was: the model was undertrained. Therefore, the focus in 2026 was on extending training time, optimizing training parameters, improving per-step training speed, and reducing gradient noise.
Specifically, the team constructed semantic stages from skill annotations in demonstrations, warm-started task and stage embeddings for new tasks from similar old tasks, and designed a 21-dimensional action space matching teleoperation control. In addition, they adopted a variance-reduced multi-sample flow matching method, paired with a simple noise-compute model that selects the number of samples per action expert for each observation without requiring ablation experiments. The final training run lasted 200,000 steps, roughly equivalent to one pass over 211 million frames of demonstration data.
A vision-language-action (VLA) model is a class of model that unifies visual perception, natural language instructions, and robot action generation into a single network. Its typical structure consists of a vision encoder (e.g., ViT), a language encoder (e.g., LLM), and an action decoder, trained on large-scale demonstration data so that the model can directly output low-level control commands based on current observations and task descriptions. The core advantage of VLA models is cross-task generalization: the same set of parameters can handle multiple instructions without designing a separate policy for each task. However, VLA models face challenges of error accumulation and insufficient planning capability in long-horizon tasks—this is precisely the context for the hybrid architecture attempt in this article.
Flow matching is a generative modeling method that generates samples by learning a continuous transformation (ordinary differential equation) from a simple noise distribution to the target data distribution. In robot policy learning, flow matching is used to model action distributions: given an observation, the model predicts a velocity field that gradually transforms noise into an action sequence. Multi-sample flow matching refers to generating multiple candidate action samples for the same observation, then selecting the optimal action according to some criterion (such as Q value or consistency). This article employs variance reduction techniques to lower the gradient noise of multi-sample estimation, and uses a noise-compute model to adaptively determine the number of samples per observation, thereby balancing computational efficiency and action quality.
Infrastructure: Shared Records and Efficient Simulation
Supporting large-scale training and evaluation is a cloud-based evaluation infrastructure. The system maintains shared records with provenance information, episode diagnostics, and recoverable simulator states, and can run six simulators simultaneously on a single GPU, faster than real-time. This efficient infrastructure enabled the team to iterate quickly and maintain continuous monitoring of model performance during training.
OmniGibson is an embodied intelligence simulation platform built on NVIDIA Isaac Sim, designed specifically for the BEHAVIOR series of challenges. It provides high-fidelity physics simulation, photorealistic rendering, and a large number of interactable household objects, supporting standardized evaluation of 100 long-horizon household tasks. A core feature of OmniGibson is support for recoverable simulator states: complete scene states can be saved and loaded during training and evaluation, facilitating diagnostics and reproducibility. The platform also integrates a teleoperation interface and demonstration data collection tools; the 21-dimensional action space design in this article is matched to teleoperation control.
Hybrid Architecture: When a Coding Agent Takes Over the Robot
Beyond the learned policy, the team routed 14 tasks to a GPT-6 Astra coding agent. The agent controls the robot through observation and motion tools, with each episode starting from a frozen copy of its practice session on training instances. On these 14 tasks, Astra achieved an average Q value of 0.639, compared to only 0.108 for the learned policy—a significant gap.
This comparison reveals the limitations of current VLA models on specific tasks, and also demonstrates the potential of coding agents in tasks requiring complex reasoning and planning. However, the team also noted that the public submission included only some Astra results, with a final score of 0.446 Q and a full-task success rate of 27.3%.
A coding agent refers to an AI system capable of writing, executing, and debugging code, typically built on large language models, with tool-calling and environment-interaction capabilities. In this article, the GPT-6 Astra coding agent obtains robot states through observation tools and issues control commands through motion tools, essentially transforming the robot control problem into a code generation and execution problem. Unlike end-to-end VLA models, the coding agent does not directly output actions but instead generates programs that call motion interfaces, thus possessing stronger symbolic reasoning and long-horizon planning capabilities, but with higher inference latency and dependence on precise tool interface design. This paradigm shows advantages in tasks requiring complex logic, yet struggles with high-frequency fine manipulation.
Insights and Outlook
The value of this work lies not only in the leaderboard ranking but also in the reusable engineering experience it provides to the embodied intelligence community. Extending training, optimizing flow matching sampling, and building efficient evaluation infrastructure—these seemingly plain improvements, when stacked together, brought significant performance gains. The exploration of hybrid architecture opens a new direction: in domains where end-to-end models are not yet fully competent, introducing coding agents with reasoning capabilities as a supplement may be a pragmatic path.
For researchers and developers focused on robot learning, the core contribution of this solution is demonstrating the scalability of a single policy model in multi-task scenarios and the critical impact of training sufficiency on performance. As the BEHAVIOR Challenge continues to be held, these experiences will drive embodied intelligence toward more general and reliable evolution.
One-line Summary
An independent team used a single VLA model plus a hybrid coding agent to place ninth among 100 tasks in the BEHAVIOR Challenge.
Paragraph Summary
This article introduces the solution of an independent team in the 2026 BEHAVIOR Challenge. The team used a single vision-language-action (VLA) model to tackle 100 long-horizon household tasks, placing ninth on the leaderboard and being the only independent entrant in the top ten. Core improvements include: extending training to 200,000 steps, optimizing flow matching sampling and gradient noise, and building cloud-based evaluation infrastructure supporting six parallel simulators on a single GPU. Additionally, the team experimented with a hybrid architecture, routing 14 tasks to a GPT-6 Astra coding agent, which achieved an average Q value of 0.639, far higher than the learned policy's 0.108. The final public submission scored 0.446 Q with a full-task success rate of 27.3%. The work demonstrates the multi-task scalability of a single policy model and the critical role of training sufficiency.
Related articles

Gemini 4 Neon Leak: A Potential Strong Contender Against Opus 5.5
A Hacker News thread suggests Gemini 4 Neon looks like an Opus 5.5 competitor. This article unpacks the product positioning behind this community signal, shifts in the top-model competitive landscape, and what it means for developers' model selection strategy.

Claude Fable 5.2's Self-Authored Short Film 'Fable' Sparks Discussion
A short film 'Fable,' credited to Claude Fable 5.2 for script, visuals, and directing, appears on Hacker News. This article analyzes the significance of full-pipeline AI creation, community reactions, and what it means for content creators.

Anthropic IPO Pricing: A Signal of AI Giants Going Public
Anthropic's IPO pricing draws attention. This article analyzes its potential impact on AI industry valuation, commercialization paths, and the competitive landscape, and highlights public-market expectations and challenges for AI companies.