CIALLO lets one robot policy reach 76.7% success on two real arms with 20 demonstrations per task

CIALLO is a method that represents the short-term intent behind robot actions in a shared form, so knowledge from one robot can be used on another. A decoder turns that intent back into commands for each robot.

After training on eight robots, it needs only one short recording, up to 10 seconds long, to adapt to a new robot. A policy built on it, called Qwen-CIALLO, averaged 76.7% success on real SO-101 and Piper arms with 20 demonstrations per task. A fully fine-tuned pi0 policy reached 70.0%.

In simulation, it scored 97.3% on the LIBERO benchmark and 52.7% on RoboTwin 2.0.