Sakana AI Method Boosts Robot Task Success from 25% to 73%

A new collaboration between the University of Tokyo and Sakana AI has produced a technique that allows robots to test and correct movements in a virtual environment before attempting them physically. Known as Scaling In-Context Imitation Learning (SAIL), the method employs a policy Vision-Language Model (VLM) to generate trajectories based on a few demonstrations, which are then reviewed by an evaluation VLM to find errors. In 6 simulation tasks, increasing the search budget from 1 to 45 candidates improved the average rate of finding a successful trajectory from 25% to 73%. The research, to be presented at the IROS 2026 conference, leverages existing foundation models to avoid training a specific policy for every new robot task. Researchers noted that while models like GPT-6 Astra can operate physical robots, a single generation of movement is often unreliable and prone to total failure from small errors. SAIL uses Monte Carlo tree search and additional computation during the test phase to refine actions in simulation before deployment.