Chelsea Finn: This is the State of the Art in Robotics

Market Trends and Industry Development - Physical intelligence companies are transitioning robotics into a GPT-like and Dolly-like era of generalist AI models [89] - Industry deployment expands into diverse robotic embodiments such as standard bimanual platforms, drones, quadcopters, surgical robots, and tractors [91] - Autonomous driving systems like Waymo achieve a milestone of 250 thousand weekly autonomous rides, demonstrating trustworthy physical AI operations [12] Technological Innovation and Performance - Physical intelligence models achieve over 90% success rate in complex tasks like making espresso [36] - Integration of reinforcement learning and human interventions delivers a 2x higher throughput by preventing dead-end trajectories [36] - Advanced multi-scale memory systems enable robots to execute non-repetitive, long-horizon tasks for 10 to 15 minutes autonomously [46] Data Strategy and Model Training - Pre-trained models like PIO7 match or outperform specialized fine-tuned models out of the box through diverse data training and metadata prompting [72][86] - Training datasets incorporate heterogeneous sources including low-quality demonstration data, policy rollouts, human videos, and web data [62][63] - Models demonstrate strong compositional generalization, enabling zero-shot execution on unseen tasks, appliances like air fryers, and different robot platforms [74][78][87]