Komatsu has begun production integration of the NVIDIA Jetson Thor edge-AI compute module into its PC490LCi-11 Intelligent Machine Control (IMC) hydraulic excavator for the 2026 model year, the first deployment of the NVIDIA platform in a production construction machine. The partnership, announced at the Caterpillar–NVIDIA alliance at MINExpo 2025, has been adapted by Komatsu into its own Smart Construction Edge architecture, which previously relied on cloud-processed drone surveys and offline digital-twin updates. By moving neural inference onto the machine itself, the PC490LCi-11 can now update its 3D design model and bucket-fill estimate in real time, without any cellular round-trip, even in deep-trench or remote-mine environments with zero connectivity.
The engineering breakthrough is the "Neural Bucket-Fill Estimator." Traditionally, IMC systems use the boom and arm cylinder pressure sensors to infer bucket payload - a noisy signal affected by machine pitch, stick friction, and joint backlash. The Jetson Thor module fuses the cylinder pressure stream with a forward-facing stereoscopic camera mounted on the cab, running a convolutional neural network trained on 4.2 million labeled bucket-load images. The model classifies the bucket fill state (underfill, optimal, overfill, spill risk) 30 times per second and adjusts the hydraulic flow to the arm cylinder to prevent the bucket from "raking out" at the end of the cut. This closed-loop fill control eliminates the typical 8% under-load cycle that forces a second pass, improving loading efficiency by 11% measured against a benchmark trained-operator baseline.
A second innovation is the "Predictive Blade-Overload Avoidance." On legacy IMC excavators, when the bucket contacts an unexpected hard layer, the system trips a hydraulic relief and halts the boom - a binary stop that forces the operator to re-plan the cut manually, wasting up to 15 seconds per event. The Jetson module runs a trained model on the IMC's machine-control GNSS and 3D design data, predicting hard-layer interfaces up to 1.5 meters ahead of the bucket by correlating as-built geology with cut resistance trends. When a hard interface is predicted, the system pre-modulates the boom flow to reduce crowd force by 40% over the last 200 mm of approach, preventing the relief trip entirely. Komatsu reports a 70% reduction in blade-overload stops and a 19% cycle-time gain in variable-geology trenching, validated on a 3.2-kilometer pipeline corridor in Queensland.
