Latent Intelligence: AI-Generated World Models for Testing Physical AI Before Real-World Deployment

This talk introduces Latent Intelligence, a startup building generative world models and AI test agents that design challenging, realistic environments to stress-test physical AI systems before real-world deployment. It covers the cost and time barriers of traditional simulation building, a live demo of autonomous failure-case generation for delivery robots, and the founders' research background from the University of Michigan.

■ The Problem with Traditional Simulation
- Manually designed test environments: M City 2.0 took about $250K and 15 months of human effort
- Scenario authoring still relies on human intuition; hard to scale for robots across varied places and tasks

■ Product & Demo
- Agent-based closed loop: a test agent proposes challenges while a world model generates and adapts the environment
- Demo of a coffee-delivery robot navigating crowds, corners, and unexpected safety-critical events
- Cost reduction to roughly $600 (about 300 GPU-hours) per environment in the current stage

■ Research, Traction & Market
- Founding team's Michigan research background and Nature cover feature
- Clients and partners: leading automakers, autonomous trucking company Bot Auto, a US DOT challenge, robot makers, and an NVIDIA Metropolis/VSS demo with the City of San Jose
- Competitive positioning against open-source entrants such as NVIDIA's Cosmos

This is useful viewing for developers of autonomous vehicles, delivery and humanoid robots, as well as simulation, safety, insurance, and regulatory professionals. Viewers will come away with a clearer picture of how generative world models can automate test-case discovery, reduce verification costs, and produce shared evidence for deployment approval — and where the field is heading next.