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

General-purpose robotics requires moving beyond task-specific training toward foundation models capable of long-term autonomy. Achieving this involves scalable reinforcement learning recipes that leverage human interventions to prevent dead-end trajectories and utilize general-purpose value functions to estimate task progress. Incorporating memory at multiple timescales—short-term video for immediate control and long-term text summaries for task tracking—enables robots to execute complex, non-repetitive workflows like cleaning a kitchen. A single, large-scale model trained on diverse, heterogeneous data exhibits compositional generalization, allowing it to perform tasks on novel robot platforms or with unfamiliar objects without additional fine-tuning. By utilizing metadata prompting, these models effectively extract utility from low-quality data, matching or exceeding the performance of specialized, fine-tuned systems and signaling a shift toward a "ChatGPT-like" era for physical intelligence in real-world deployments.