Robot learning breaks compute dogma: data trumps model size
xiaomi's robot learning model proves that increasing training data improves performance significantly more than increasing model size, signaling a shift away from compute-heavy scaling laws for embodied ai.
xiaomi introduced robotics-1, a foundation model demonstrating that 'more data beats bigger models' for robot control, a finding confirmed by benchmark tests where success rates jumped from 25% to 75% with more training. to solve the data bottleneck, the team collected 100,000 hours of manipulation data using portable handheld grippers across 1,700 environments, rather than expensive physical robots. unlike llms, scaling research for robotics indicates that compute resources are best spent on data collection rather than parameter expansion. this approach enabled the robot to perform complex tasks like autonomously packing a suitcase, marking a potential inflection point for how embodied ai systems are developed.