Sakana AI introduces SAIL for test-time scaled robot trajectories

Sakana AI and the University of Tokyo are introducing Scaling In-Context Imitation Learning (SAIL), to be presented at IROS 2026. A policy vision-language model generates robot trajectories from a few successful demonstrations without changing the model itself.

Each candidate is tested in a simulator. An evaluation vision-language model reviews the resulting video and identifies where progress stalled, and the policy model uses that feedback to revise the trajectory. Monte Carlo tree search explores alternatives while refining promising candidates, and only the selected trajectory is sent to the physical robot.

Across six simulated manipulation tasks, increasing the search budget from one candidate to 45 raised the average rate of finding a successful trajectory from 25% to 73%. Sakana AI also evaluated SAIL on a physical robot and says the results suggest that additional computation can help the model test and refine proposed actions in simulation. The SAIL blog and paper give further details.