Sergey Levine: Humanoid Robotics Results, Chinese Labs & Future Timelines

Humanoid robotics is transitioning from specialized, hand-designed controllers to general-purpose foundation models, mirroring the scaling success seen in large language models. While the industry currently focuses on establishing fundamental technologies rather than predictable scaling, the path to deployment relies on creating positive feedback loops where autonomous experience improves model robustness. Generalization remains the primary hurdle, requiring robots to navigate unstructured environments beyond the factory floor. By grounding models in physical experience and utilizing intermediate "thinking" steps—such as predicting future states or images—researchers can transfer skills across diverse robot morphologies. This shift toward learning-based systems, supported by broader data sharing and reliable, low-cost hardware, is essential for moving robotics from controlled, staged demonstrations to practical, real-world utility within the next decade.