Antioch lands Amazon Ring as a customer to validate its hybrid simulation approach for training physical AI systems

Antioch

Harry Mellsop co-founded Antioch after working on physical AI at Stanford and on Tesla's autopilot team. The company's thesis is straightforward: industries that have moved fastest on AI are the ones that achieved recursive self-improvement, and the physical world hasn't got there yet.

The gap, Mellsop argues, is simulation fidelity. Classical simulation engines, the video-game-style physics environments most teams use today, require engineers to manually encode every real-world variable. Miss the wind, add the wind, repeat indefinitely. End-to-end learned world models are promising but data-hungry, and physical AI deployments don't yet generate the volume of training data that, say, software or language tasks do.

“Our cofounding team at Antioch all met at Stanford working on physical AI... We announced our partnership with Amazon, and in particular the Ring team at Amazon — a smart security device. Any system that has a hardware, software, machine learning component where testing in the real world is really difficult and really expensive — there are tens of thousands of companies here.”

Antioch's answer is a hybrid: use classical simulation where it holds, then learn the gaps from real-world data. That learned correction feeds back into a better simulator, which trains a better physical system, which generates richer real-world data. The loop is the product.

Amazon Ring is the first publicly named enterprise customer. Ring's smart security devices sit squarely in Antioch's target profile: hardware-software-ML systems where real-world testing is expensive and edge cases are hard to enumerate in advance. Mellsop uses the win to illustrate that the addressable market extends well beyond humanoid robots and autonomous vehicles. Robotic vacuum cleaners, camera drones, and consumer security devices all face the same sim-to-real problem.

On the autonomous driving timeline debate, Mellsop is more optimistic than the incumbents. Tesla and Waymo work, he argues, because they built a data flywheel at scale, with Tesla's consumer fleet essentially crowdsourcing training data from paying customers. The bottleneck for robotics is the same flywheel in an earlier, thinner form: deployments are nascent, so real-world data is scarce, which is exactly why sample-efficient simulation matters now.

Antioch's bet is that a company which can bootstrap that flywheel through simulation, before deployments reach meaningful scale, sits at the critical chokepoint as the physical AI market matures.