From Learning Humans to Learning Robot Peers: A New Path for Scalable Embodied I

A joint research effort by Tsinghua University, Wuwen Xinqiong, and Zhiyuan proposes Collective Experience Learning (CEL), enabling robots to learn from each other's experiences rather than relying solely on human demonstrations. Validated on a 96-robot cluster, CEL achieved task success rates exceeding human benchmarks in complex scenarios like mobile phone packaging. The study demonstrates the scalability of the approach, with performance improving as the number of robots increases, and highlights the integration of robust system design and algorithmic innovation to support continuous learning in real-world environments.

A joint research effort by Tsinghua University, Wuwen Xinqiong, and Zhiyuan proposes Collective Experience Learning (CEL), allowing robots to learn from each other's successes in a 96-robot cluster. The method demonstrates improved task success rates, with robots achieving over 97.5% success in various real-world tasks, including complex industrial operations. The study highlights the potential for scalable, self-improving robotic systems through shared experience and continuous learning.