96 robots trained to learn from each other hit 97.5% success

Researchers from Tsinghua University, Infinigence AI and AgiBot proposed Collective Experience Learning, in which robots learn from peers' successes rather than mainly human demonstrations. On a cluster of 96 real robots, success across five tasks reached 97.5%, and phone packaging rose from 79.2% to 98.3%.

• Two untrained robots ran 600 trials with 15 failures — 97.5% success
• Phone packaging success rose from 79.2% to 98.3%, 2 failures in 120 runs
• Scaling from 16 to 96 robots lifted success from 86.7% to 98.3%
• Cluster collected 9,834 episodes; 30 faults cost only 1.3% of robot time