A team of independent researchers, including Ilia Larchenko, Gleb Zarin, Atindra Nair, Keerti Manney, Vitaly Kleban, Mirko Kiefer, and Anjana Yodaiken, placed ninth in the 2026 BEHAVIOR Challenge, a competition evaluating robot policies on long-duration household activities in the OmniGibson simulator. Their solution utilized a single vision-language-action (VLA) model, a refinement of their 2025 winning entry, to address 100 distinct tasks. The model attained a Q-score of 0.392 across all 100 tasks.
The BEHAVIOR Challenge, built on NVIDIA Isaac Sim, assesses embodied AI agents on their ability to perform activities such as turning on a radio or cooking a meal. The 2026 iteration expanded the task set from 50 to 100 activities compared to the 2025 challenge. Each task in the 2026 challenge included 200 teleoperated demonstrations, totaling 20,000 episodes and 211 million frames of data. The robot used in the simulation, a Galaxea R1Pro, features a holonomic base, a four-joint torso, and two seven-degrees-of-freedom arms with parallel grippers, observed through head and wrist cameras, along with proprioception.
The independent team's approach centered on a single VLA model checkpoint designed to handle all 100 tasks, a departure from their 2025 solution which used four checkpoints, each fine-tuned for a subset of 50 tasks. The 2026 model was warm-started from the 2025 checkpoint that contributed most to their previous year's score. The researchers' primary focus for the 2026 challenge was extended training, optimization of training parameters, and reduction of gradient noise. This involved a migration from 50 to 100 tasks, with semantic stages built from demonstration skill annotations, and task and stage embeddings for new tasks initialized from similar older ones. They also implemented a 21-dimensional action space to match teleoperation controls.
The team's Q-score on the 50 tasks shared with the 2025 challenge increased from 0.260 to 0.371. The Q-score metric, used for ranking policies in the BEHAVIOR Challenge, measures the fraction of completed sub-goals and awards partial credit for meaningful progress, offering a more nuanced comparison than a binary success rate. The first-place entry in the 2026 BEHAVIOR Challenge, "HiPhys-RSI," achieved a Q-score of 0.683 and a 48.2% success rate across all 100 tasks.
The OmniGibson simulator, which hosts the BEHAVIOR Challenge, provides realistic physics and rendering for rigid bodies, deformable materials, and liquids across 50 detailed indoor scenes with over 9,000 unique object models. The BEHAVIOR-1K benchmark, which includes 1,000 everyday activities, is designed to test embodied AI agents on human-centered tasks such as cleaning, cooking, and organizing. The 2026 challenge offered a total prize pool of $11,000, with $5,000 for first place.