Tsinghua, Infinigence and AgiBot Validate Collective Experience Learning on 96 Robots

The proposed Collective Experience Learning framework shifts robot learning beyond human demonstration. In the CEL setup, each robot’s successful task trajectories and recovery strategies are recorded and shared with the collective. When a robot encounters a difficult state, it can retrieve relevant past experience from peers and use it to adjust its behavior, rather than relying solely on fresh human teleoperation or retraining.

The researchers from Tsinghua University, Infinigence AI and AgiBot evaluated CEL in a cluster of 96 real robots. The large-scale trial tested whether experience accumulated by one robot could improve the performance of others in continuous practice. According to the team, the robots could continually refine their policies by sharing and accumulating experience, demonstrating that collective learning can operate on physical robot fleets rather than only in simulation.

The team positions CEL as a technical route toward autonomous robot evolution and scalable embodied AI deployment. By reducing dependence on human demonstrations and enabling robots to learn from one another, the approach could help fleets adapt faster in real-world environments. The validation across 96 real robots suggests that collective experience sharing may become a practical mechanism for improving robot capabilities at scale, though further research is needed to assess generality and long-term reliability.