Antioch combines classical simulation with learned models for physical AI

Antioch uses a hybrid strategy to train physical autonomous systems: classical, high-fidelity simulation is applied where it works, while learned models fill gaps between simulation and reality. The company says both classical simulation and end-to-end world-model approaches currently face a substantial sim-to-real gap; classical systems can require continual additions for factors such as wind, while world models lack the large volumes of physical-world data needed for training.

Antioch believes world models will ultimately be the right approach for physical AI and that the field is moving in that direction, but says the technology is not yet fully ready for broad use. Its current approach is intended to help real companies transition toward world-model-based systems over time, alongside the company’s stated goal of enabling recursive self-improvement for physical autonomous systems. Antioch’s co-founding team met at Stanford, and one co-founder previously worked on Tesla’s Autopilot team.