An interview with CoreWeave Physical AI SVP Richard Ahlfeld on AI models failing real-world checks, the roles of synthetic data and physical tests, and more

Physical AI models are now being deployed with dedicated field engineering support to ensure they perform in real-world conditions.

CoreWeave has launched a Physical AI Field Engineering service, aiming to address failures in AI models when encountering real-world scenarios not present in training data. This service leverages engineers with expertise in automotive, aerospace, and mechanical fields to build AI models using existing customer data from sources like test benches and sensor telemetry. The initiative stems from CoreWeave's acquisition of Monolith, an engineering AI company, and has reportedly been applied to over 100 projects. The company emphasizes that the most common AI failures are not due to flawed models but rather incomplete or unrepresentative training datasets. CoreWeave's approach, which includes utilizing tools like Weights & Biases for experiment tracking and ARIA for research, has been employed with clients such as Nissan and the Aston Martin Aramco Formula One Team. The service focuses on augmenting data and refining models based on physical tests and simulations to improve reliability for critical applications.

The core challenge for physical AI is not model architecture or compute power, but the inadequacy of training data to capture rare but critical real-world events. CoreWeave's new Field Engineering service addresses this by embedding domain experts who can identify and source the right data, moving beyond simply collecting more data to actively seeking crucial edge cases. This signals a shift towards more hands-on, expert-driven AI development for physical systems, acknowledging that understanding the physics and the data generation process is as vital as the modeling itself.

This physical AI approach, proven across 100+ projects and used with clients like Aston Martin, demonstrates a critical need for validation beyond simulation, especially for soft materials or nuanced environmental conditions. The reliance on domain engineers to approve agent actions on hardware underscores the current limitations of full AI autonomy in safety-critical physical applications. The success of this model suggests a growing market for specialized AI engineering services that bridge the gap between theoretical models and practical, reliable deployment in complex physical environments.