RSI Gains Traction: Robot Deployment Is Becoming the Starting Point for the Next Round of Training

The article analyzes the rise of recursive self-improvement (RSI) in embodied intelligence, noting that Anthropic's definition requires AI to autonomously develop successor systems, while most current systems remain bounded self-improvement. 1X established the World Model Lab to close the loop from pretraining to real-world data; Motus2 uses three interfaces—Policy, Simulator, and Evaluator—to achieve bounded self-evolution; and Pudu's PuduFM 1.0 relies on more than 130,000 deployed robots to build a data flywheel.