Robotics
Robots Run Better When AI Moves Off Their Bodies
Microsoft research challenges the standard practice of putting AI compute directly on robots, showing that offloading inference to remote servers boosts performance, extends battery life, and slashes costs.
Microsoft Research Blog
The robot-building industry has long assumed that AI inference must live on the robot itself. Microsoft Research questions that premise. In a systematic study of robotics workloads, they found that shipping inference tasks to remote GPUs—whether on the edge or in the cloud—can deliver faster response times, sharper task success, longer battery hours, and lower hardware costs.
Why Onboard AI Holds Robots Back#
Most robots today carry their own GPU, wired directly to control decision-making in real time. This approach seems logical until you consider its drawbacks. GPUs drain power quickly, eat into battery life, add weight and expense, and lock robots into whatever hardware generation is already installed. As AI models grow larger and more capable, squeezing them into robots' cramped compute budgets becomes ever harder.
The Measurement Study#
Microsoft researchers studied mobile manipulation tasks—the kind where a robot might navigate to a kitchen, spot trash, and deposit it in a bin. They tested representative AI models for three core capabilities: semantic mapping and planning, obstacle navigation, and physical manipulation. Performance varied wildly depending on what GPU the models ran on.
Smaller onboard options like the Jetson Nano couldn't even fit the full manipulation stack. Lighter GPUs slowed semantic mapping by as much as 383% compared to an A100. Navigation lost 30% of its obstacle-detection speed. Vision-language models (VLAs) dropped their accuracy by half. By contrast, when the same workloads ran on more powerful remote GPUs, robot success rates climbed—especially in demanding tasks like object handover between manipulators.
Battery Life Transformed#
Beyond pure performance, the energy argument is stark. When researchers replaced an onboard GPU with a lightweight Raspberry Pi and sent all inference data to a remote processor, battery life extended by over 100% in some cases—adding hours of operation on a single charge. For a Stretch-3 robot, swapping a power-hungry Jetson Thor for offloaded compute nearly doubled its time between recharges.
New Tooling for Deployment#
Microsoft has released an expansion to its Physical AI Toolchain (opens in new tab) that automates inference offloading. The system uses Kubernetes to containerize and distribute robotics AI workloads across the robot's onboard hardware, nearby edge servers, and the cloud. Developers can declare what to offload, and the toolset handles the rest—ideal for teams using and infrastructure.
