Caterpillar and CoreWeave are collaborating to accelerate the development of physical AI for construction machinery. The partnership aims to reduce the time required to train autonomous equipment, shortening feedback loops from months or weeks down to a single workday.
The initiative leverages CoreWeave's specialized infrastructure and its new Physical AI Field Engineering service. This service embeds CoreWeave engineers directly with Caterpillar's domain experts to optimize the training process for autonomous excavators and other heavy equipment.
Training these systems requires massive data processing, including telemetry, vision, and performance metrics. Caterpillar currently manages 18 petabytes of federated data, with individual machines generating terabytes of information daily through sensors like LiDAR and cameras.
The companies utilize AI models in collaboration with Nvidia to automate the annotation and labeling of field data. This integration allows for rapid simulation and reinforcement learning, enabling machines to adapt to the unstructured environments typical of construction sites.
