Offloading Physical AI Inference from the Robot to Edge or Cloud GPUs: A Systematic Measurement Study of Mobile Manipulation Workloads
This work systematically measures mobile robotic manipulation workloads (semantic mapping and planning, navigation, and manipulation) across onboard, edge, and cloud GPU configurations, reporting that offloading inference off the robot improves task performance and battery lifetime, and releases Kubernetes-based automatic offloading tooling as a new capability in the Physical AI Toolchain.