Every lesson so far has trained a policy once and frozen it — clone the behavior, then deploy. That is not how the frontier operates anymore. Physical Intelligence's π*0.6 runs continuous on-robot reinforcement learning through a method called RECAP: in one evaluation it kept making espresso from 5:30am to 11:30pm — about 18 hours straight — then folded 50 novel laundry items in a home it had never seen, both above 90% success (arXiv:2511.14759). AgiBot's LWD closes the same loop at fleet scale: 16 dual-arm robots learning while deployed, sharing autonomous experience across eight real-world tasks, landing one generalist policy at 95% average success (arXiv:2605.00416). Read together, they refute the frozen-checkpoint model this course has taught so far: deploy, collect autonomous experience, update, redeploy — the flywheel — is how 2026's best policies are made, not a research curiosity.
But continuous interaction is now table stakes; both of the above accept an update because it scored well on average, not because anything certifies it won't make the worst case worse. The Institute's answer — a target, not a shipped system — is to make one scalar do both jobs: an energy function that is simultaneously what the flywheel descends and a formal stability certificate, so a correction is rejected before it is committed if it raises the worst case, not only if it looks bad on average. The bench below builds a toy version of exactly that gate.