Unified policy for a hundred household tasks: an independent team breaks into the top ranks of the 2026 Behaviour Challenge
Listen to this article
Read by Anchor
Does training robots to perform a hundred complex household tasks require multi-billion research budgets and computing centers monopolised by major global companies? A new research paper published by Eliya Lartshinkov and his team, which includes Gleb Zarin, Attendra Nair, Kerti Mani, Vitali Kliban, Merko Keifer, and Anjana Yodai-Ken, shows that engineering efficiency can break this monopoly; the team finished ninth in the final ranking of the 2026 Behaviour Challenge, being the only independent team that managed to enter the top ten amid fierce competition within the Omni-Gibson simulation environment.
The team’s core solution relied on a unified policy and a single software checkpoint for the vision, language, and motion model, built as an extension of their first-place winning solution from 2025 and based on the Bay 0.5 model.A review of the previous cycle’s results showed that the model suffered from a lack of training rather than a structural flaw.This prompted the 2026 researchers to double the training time and improve their parameters, as well as accelerate computational steps and reduce the noise of the mathematical gradient across a hundred full tasks.
Upgrading from fifty to a hundred tasks required a semantic split of work stages based on skill descriptions recorded in the experimental offers, with directional embeddings prepared for the new tasks and stages based on similar previous tasks, and the application of a 21-dimensional motion space that matches remote-control tools exactly. The team also used a low-variance multi-sample flow-matching technique with a simplified model that balances noise and computation to determine the number of expert motion samples for each visual scene, through a training path that spanned two hundred thousand steps and consumed roughly one pass through 211 million recorded visual frames.
These improvements raised the quality index for the fifty shared tasks with the 2025 cycle from 0.260 to 0.371, while the index across all one hundred tasks reached 0.392. To complete the test cycles without high costs, the team built a cloud-based evaluation infrastructure that preserves source records, round diagnostics, and recoverable simulation states, and succeeded in running six simultaneous simulation environments on a single GPU at a speed exceeding real time.
The most significant technical shift was the team’s recognition of the limits of a purely neural policy; fourteen complex tasks were directed to a software agent powered by the GPT-6-Astra model that controls the robot via monitoring and motion tools, starting each round from a frozen copy of previous training sessions.The Astra agent achieved a decisive jump in the quality index for those tasks, recording 0.639 points versus 0.108 points for the learned policy.This raised the overall score for the announced participation partially to 0.446 points with a full task success rate of 27.3 percent.
These results carry direct implications for developers and engineering teams in Saudi Arabia, the United Arab Emirates, Egypt, and Jordan; many robotic and autonomous systems engineers hesitate to adopt embodied AI models, believing that completing sequential tasks requires a cloud infrastructure only large Western labs possess. The experiment shows that reducing simulation time by running six environments on a single GPU allows regional teams to test their ideas at an affordable cost, and the hybrid distribution of work between a motion neural model and a smart software agent provides a clear roadmap for building logistics automation and facility management solutions without waiting for closed proprietary models.
The team made the code, checkpoints, and project page available to everyone, setting a practical benchmark that demonstrates that excellence in robotic task processing stems from sample tuning, training stability, and intelligent software integration, not merely from consuming massive computing capacities without a refined architecture.