Intel / article

SAIL: Scaling In-Context Imitation Learning

Source: unknown
Author: Sakana AI
Date: Published Sep 27, 2026

Context

Sakana AI's SAIL tests and revises proposed robot movements in simulation before sending a selected movement sequence to the physical robot. A vision- generates candidates from successful demonstrations; another pass examines simulated video and returns feedback about stalled progress. A search process explores alternatives and refines promising candidates without changing the model's weights. The project article describes this loop.

Across six simulated manipulation tasks, increasing the budget from one candidate to 45 raised the average rate of finding a successful trajectory from 25% to 73%. Success here means passing the simulator's task check. Physical testing was much smaller: five successes in six trials of one block-placement task, leaving broader real-world transfer unresolved.

Key takeaways · AI-distilled

Sakana AI and the University of Tokyo note a single VLM generation does not necessarily yield a reliable robot trajectory: output depends on , and one small error in a movement target can fail the whole task.
SAIL runs each candidate trajectory in a simulator, has a separate evaluation VLM watch the video to find where progress stalled, and uses Monte Carlo tree search to explore and refine alternatives.
Across six simulated manipulation tasks, raising the search budget from one candidate to 45 lifted the average rate of finding a successful trajectory from 25% to 73%.
Only the selected trajectory reaches the physical robot; the authors say how far these simulated gains carry over to real hardware is still an open question.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.

Source link unavailable.unknown