Classical robot navigation stacks are reliable but often need re-engineering for each new site or robot. Learning-based navigation shifts the work to data generation and training, but whether policies adapted in simulation stay reliable on real industrial robots has had little systematic testing. The authors benchmark NVIDIA's X-Mobility and COMPASS workflow on Idealworks robots, adding two embodiments and a digital twin of the Idealworks R&D center, then deploying policies on an onboard Jetson AGX Xavier.
Training the COMPASS policy in the digital twin with Cosmos sim-to-real augmentation raised real-world success from 65.3% to 91.7% across four scenarios, compared with simulation-only training. It also nearly halved weighted travel time and brought performance within 8.3 points of the production navigation stack, fully onboard. With an integrated global planner, the augmented policy reached far, occluded goals in every trial.
The paper measures the sim-to-real gap by running static scenarios both in a digital-twin replica and the real facility, and contrasts Cosmos-augmented data against COMPASS trained only on NVIDIA's environments. The authors note that learning-based navigation still trails the production stack on absolute success rate, but the benchmark provides a systematic baseline for closing that gap in industrial intralogistics.