AI news · October 8, 2026
Caterpillar and CoreWeave shorten physical-AI training loops
Caterpillar reportedly manages 18PB of federated machine data and wants simulation and reinforcement learning to cut physical-AI feedback to hours. The report describes an industrial workflow, not a ready-made product.
- reported Caterpillar federated data
- 18PB
- reported data from one machine per day
- terabytes
- target feedback time instead of weeks or months
- hours
SiliconANGLE reported on October 6 that Caterpillar and CoreWeave are working on a shorter learning loop for physical AI. The report says Caterpillar has about 18 petabytes of federated data and that one machine can produce terabytes of LiDAR, camera, control, and performance data per day. The described workflow uses NVIDIA models to annotate incoming data, simulation to run an excavator scenario many times, and reinforcement learning to improve the result. The goal is to move from months or weeks of feedback to hours within the workday.
This is a useful pattern for robotics and industrial systems because the hard part is not only the model. Data capture, labeling, simulation, and safe testing must run as one loop. The story is a secondary report from a Fully Connected event, and its performance claims come from company representatives. Smaller teams can copy the method on one simulated task before buying a large cluster: measure data preparation time, simulation throughput, transfer to hardware, and whether the learned policy survives real conditions.
What you can do with it
Use the CoreWeave event description as a starting point for one simulated physical task. Record the time spent collecting data, labeling it, running simulation, training, and testing on hardware. Keep a real-world holdout set so faster iteration does not hide a policy that fails outside the simulator.
Our take
The important product is the feedback loop, not a single robotics model. If simulation and data preparation take weeks, more model capacity will not rescue the schedule. The report is early and company-led, but it gives builders a concrete checklist for testing physical AI with a smaller, safer experiment.
Links Report Event details
Source: SiliconANGLE ↗ — Made With Models writes the brief; the reporting is theirs.