Sakana AI unveils SAIL, scaling in-context imitation learning for robots without retraining
SakanaAILabs · x · 2026-09-28
Sakana AI and the University of Tokyo introduce SAIL (Scaling In-Context Imitation Learning), to be presented at IROS2026. Instead of collecting demonstrations and training a new policy for every task, SAIL tries to extract the robotics knowledge foundation models already hold from images, text, and robot-related data — without modifying the model itself. The authors note GPT-6 Astra has been shown to operate physical robots, and prior work has LLMs/VLMs generating whole movement sequences from a few demos, but foundation model outputs don't reliably translate into robot control; SAIL targets exactly that gap.
More from Embodied
- Watch Opus design a 3D asset, print it, and have a robot arm retrieve it — burny_tech · 2026-09-28
- Robot spotted collecting embodied AI data at a night market — xiaosun86 · 2026-09-28
- CMU's DeformX trains robots to whip ropes in sim — UR5e knocks apple off a head with 0cm error — DJiafei · 2026-09-28
- Figure's Helix 2.5 results make this founder rethink why humanoids look human — yongqianme · 2026-09-28
- Sharpa Robotics wows with dexterous robot hand demo open to public challenge — jonstephens85 · 2026-09-28
- Jev seen as fit for high-level robotics orchestration: fast schema beats LLM lag — ai · 2026-09-28