Simulation in the loop: the Everyday Robots patent behind our world loop
US11938638B2 is often described as a sim-to-real patent. The title undersells it. What it actually describes is using a live simulation as an active component of controlling a physical robot, not just a training ground you leave behind at deployment. The named inventors are Matthew Bennice, Paul Bechard, Yunfei Bai, Tigran Gasparian, Brent Austin, and Andreas Christiansen, from the Google X and Everyday Robots effort. Two of them, Matt and Paul, are Cereal founders.
The problem it solves
Traditional robotics works in three steps: train a policy in simulation, deploy it to a real robot, and hope reality stays close enough to the training distribution. Once the robot ships, the simulator is out of the loop. The patent argues that is backwards. Reality changes constantly, and a policy frozen at training time cannot ask what will happen before it acts.
The invention: rehearse before every action
Instead of executing immediately, the system reads the robot's current state, updates a digital twin until it matches reality, spins up many hypothetical futures, tests candidate actions inside the simulator, picks the best outcome, sends only that action to the real robot, and repeats. The simulator becomes a predictive planning engine that participates in every control cycle. The robot stops asking what should I do and starts asking what would happen if I did this.
Three properties fall out of that loop. The twin stays continuously synchronized with the world: object locations, robot pose, arm configuration, sensor state. Futures are evaluated in parallel, which is model-predictive control fused with modern simulation. And replanning is closed-loop: when an object slips, the old world updates, the futures regenerate, and the plan changes, instead of the task simply failing. Learning still matters, but it sits inside the loop: models propose actions, simulation evaluates them, the planner chooses.
Why Everyday Robots cared
The Everyday Robots program's hardest lesson was that robots fail in homes and offices because reality refuses to hold still. Rather than chasing a perfect policy, this architecture lets a robot continuously ask which action is most likely to work given what it sees right now. Every movement is rehearsed before it is performed, which is what safety, robustness, and recovery from surprise actually require in unstructured space.
Where Cereal builds above it
The patent answers one question: how should a robot decide its next physical action. Cereal's thesis lives one layer up: how does the same intelligence persist across many robots. Persistent household memory, permissions, the embodiment contract, and the attested audit trail are not covered by simulation-guided execution, and simulation-guided execution is not replaced by them. The two are complementary, and our world-loop architecture, where the twin verifies every plan before a body moves, is the direct descendant of this thinking. The patent is assigned to the inventors' prior employer; what Cereal carries forward is the people who invented it and the conviction that the simulator belongs inside the deployed system, with the persistent layer built on top.