Key facts
- Laplace Intelligence is building automated red-teaming for robotics and physical AI.
- The company uses high-fidelity simulations reconstructed from the real world, paired with AI agents to find failure points.
- Laplace's simulation environment is a post-trained version of Nvidia's Cosmos platform.
- Developers can use language prompts to alter simulation settings in real time, incorporating real-time traffic data.
- Laplace's AI agents are designed to automatically create corner cases by analyzing information from news and collision reports.
- The startup is generating revenue from autonomous driving use cases and has customers including Isuzu, Bot Auto, and Nvidia.
Laplace Intelligence is developing a system for automated red-teaming, a process of stress-testing AI by deliberately trying to make it fail, specifically for robotics and self-driving cars. The company, founded a year ago by Howie Sun and Henry Liu out of the University of Michigan, aims to ensure the safety of physical AI before its deployment.
CEO Howie Sun stated that the unpredictability of AI agents, as seen in incidents like the OpenAI-Hugging Face event, highlights the need for robust testing in the physical world. Laplace's approach uses high-fidelity simulations, built upon Nvidia's Cosmos platform and grounded in real-world data, to recreate environments like highways and pedestrian crossings. These simulations can be dynamically altered using language prompts and real-time traffic information from providers like TomTom.
The core innovation lies in Laplace's AI agents, which are designed to automatically identify and create challenging corner cases by analyzing various data sources, including news and collision reports. This moves the process of identifying edge cases further away from human intervention. The company's academic roots are evident in its methodology, which is based on research published in the journal Nature.
Laplace Intelligence is currently generating revenue from project-based contracts, primarily from autonomous driving use cases, with clients such as Japanese trucking manufacturer Isuzu and Texas-based autonomous trucking firm Bot Auto. Nvidia is also a customer, and Laplace is collaborating on a smart city pilot in San Jose. The startup is also exploring partnerships for humanoid robotics applications in warehouses. The company's near-term goal is to raise funds to scale its operations and compete with rivals like Applied Intuition.