Offloaded inference for real-world physical AI robotics

Microsoft Research systematically measured mobile manipulation robot workloads and found that offloading physical AI inference from onboard GPUs to edge or cloud GPUs can improve task success, support larger models and extend battery life.