In the race to build better AI models, most of the attention has gone to pre-training. But in robotics, we believe post-training is just as critical. Pre-training can get you impressive capabilities. Post-training is what closes the gap between a demo and a deployment—between a video and actual dollars. In fact, the real frontier in robotics may increasingly lie in post-training. Today, we’re introducing a new approach to post-train robotic policies using self-play. Inspired by the original self-play work at DeepMind, as well as our own work on robust adversarial reinforcement learning, we train policies in simulation to help emerge behaviors and make them robust before they ever reach the real world. This is an early step (and on sports), but we believe it points toward something much bigger: bringing some of the original ideas that made reinforcement learning so powerful back into the physical AI stack.

Personally, working on the S1 release with the team has been one of the most satisfying accomplishments of this year. I’ve spent countless late nights discussing some of my favorite topics—like affordances and how we can build systems that truly generalize. We’ve celebrated every step change in generalization: from new object arrangements, to different objects, to where we are today—being able to perform completely unseen tasks. When I started working in robotics, I never imagined we’d reach a point where a robot could perform a task it had never been trained on, simply through prompting. And we’re proud to say that no other research group has yet demonstrated generalization to completely unseen tasks, or to long-horizon tasks like making pancakes.