Sakana AI's SAIL lifts robot trajectory success to 73%

Sakana AI, together with the University of Tokyo, presented SAIL (Scaling In-Context Imitation Learning), a method for making robot trajectories more reliable by searching at inference time rather than retraining the policy. The paper is due to appear at IROS 2026 and is available as arXiv 2603.08269.

The motivating observation is that a foundation model can produce a plausible robot trajectory and still fail, because a small error in one movement target fails the entire task. SAIL adds a loop around the policy model. The policy VLM generates candidate trajectories conditioned on a handful of successful demonstrations, those are run in a simulator, an evaluation VLM reviews the resulting video and identifies where progress stalled, and the policy VLM then revises. MCTS explores the alternatives, and only the winning trajectory is sent to the real robot.

The headline result: across six manipulation tasks in simulation, raising the search budget from 1 to 45 candidates lifted average success from 25% to 73%. The authors also evaluated the method on a physical robot.

Two caveats before this gets read as a robotics result. The 25% to 73% figure is a simulation number, and the physical robot evaluation is presented qualitatively rather than as a comparable success rate. The work is better framed as a step toward understanding how much existing models can already do when they are allowed to see and correct their own failures, which is a cheaper path than training a better policy.

(Original source: Sakana AI)