Tsinghua, Infinigence and AgiBot Propose Collective Experience Learning for Robots

Tsinghua University, together with Infinigence AI and AgiBot, has proposed a Collective Experience Learning (CEL) paradigm for robots. The framework is intended to complement human demonstrations by letting robots draw on successful experiences from other members of the same fleet. Successful strategies and experience are shared and accumulated across the group, so individual robots can reuse what has already worked elsewhere instead of learning only from scratch.

By continuously practicing and pooling experience, the robot group can improve its capabilities collectively. The researchers validated CEL at scale on a cluster of 96 real robots, showing that the paradigm can operate in real-world settings rather than only in simulation. The approach treats each robot’s successful rollouts as reusable training signals for the collective, creating a feedback loop in which gains from one unit can benefit others.

According to the team, the result offers a new technical path for autonomous robot evolution and large-scale deployment. It could also reduce reliance on manual demonstration data and raise the overall learning efficiency of robot swarms. Such cross-robot experience sharing may help fleets adapt faster in changing environments and scale beyond settings where extensive human demonstration is available.