Put the robot's larger brain outside its body
Microsoft Research published measurements and tooling for offloading physical-AI inference to on-premises edge or cloud GPUs on September 23. The team reports that added latency from smaller onboard GPUs was enough to cut VLA-model accuracy by 50% in its experiments, while more capable remote GPUs improved success in object-handover and navigation tasks.
For power, the researchers replaced a heavy onboard GPU with a lightweight Raspberry Pi 5 and sent inference elsewhere, reporting more than 100% longer battery life for a Stretch-3 robot in one setup. A Kubernetes-based Physical AI Toolchain containerizes workloads, places them across robot, edge and cloud resources, and integrates ROS2, LeRobot and simulators.
The network becomes a component and a failure mode
Offloading reduces onboard weight and power but adds latency, bandwidth dependence, outages and cloud cost. Emergency control cannot depend on a network round trip. Results also come from selected robots, models and networks; reported GPU power differences and accuracy changes do not automatically transfer to every platform.
Deployments should test local safety control, latency limits, disconnected behavior, video-data protection, edge failover and end-to-end task success together.