Actuate 26 | Bringing Physical Intelligence to the Real World by Chelsea Finn

In this talk

1. 0:12

   From demos to reliable service

   Finn distinguishes impressive robot demonstrations from systems that deliver value through dependable, extended operation. Her office espresso example requires precise portafilter handling, careful movement with liquid, and good timing; the target was at least 90% success. Reaching that reliability takes repeated improvements to training data and methods, which Finn argues robots should eventually help drive themselves.

2. 3:51

   Faster reinforcement learning for robots

   Finn adapts ideas such as PPO and GRPO to robot learning, where a million one-minute attempts would amount to about 700 robot days. Her approach uses teleoperation to help robots recover from dead ends, such as separating two cardboard boxes before folding, and trains a general value model to estimate progress toward success. She reports that this method supported an office coffee service that ran for about 13 hours, more than 90% espresso-task success, and a 2× throughput increase after reinforcement-learning post-training.

3. 12:55

   A general model that works out of the box

   Finn describes a goal of moving from task-specific robot training toward models that can handle tasks without fine-tuning and combine skills in new situations. Her π0.7 model uses varied robot demonstrations, robot rollouts, human videos, and web data, with detailed task context, episode metadata, and predicted subgoals to guide actions. In the reported evaluations, the pretrained model matched or exceeded earlier fine-tuned specialists on several tasks.

4. 22:57

   Testing transfer and data diversity

   Finn tests generalization with a rare air-fryer task and shirt folding on a BARM UR5 that had no laundry-folding examples in its training data. The robot managed the shirt-folding task despite differences in size and joint configuration from the robot used for demonstrations, though it missed some grasps. Ablations showed that removing diverse data hurt held-out task performance, while metadata conditioning helped the model benefit from lower-quality data.

5. 28:31

   Progress toward real-world deployment

   Finn says π0.7 showed signs of compositional generalization and matched or surpassed fine-tuned specialists in out-of-the-box evaluations. She also describes partner deployments of Physical Intelligence models on different robot platforms for warehouse logistics, laundry folding, and excavator operation. She stresses that robots remain slower than people and still make mistakes, so substantially more work is needed for human-level deployment.