Where should a robot's brain live? Microsoft ran the experiment and found robots think faster and more accurately when the AI runs on GPUs outside the robot.
Microsoft Research tested "inference offload": running a robot's AI on nearby edge or cloud GPUs instead of the GPU mounted on the robot. On a task where the robot moves while grasping objects, the small onboard GPU made mapping and motion planning up to 383% slower, and halved the accuracy of the vision-language-action (VLA) model that turns camera input and instructions into movement. Microsoft also added Kubernetes-based routing of inference across robot, edge and cloud to its open-source Physical AI Toolchain.
Putting large AI models on a robot runs into two walls: how big a GPU the body can carry, and how long the battery lasts. Moving the work off-board makes the robot lighter and extends battery life. What this study adds is scale — the gap is not marginal, it is large enough to halve accuracy. Offloading brings its own problem, though: what the robot does when the connection drops. That question is not addressed in what we could confirm.
Whether the same setup shows up in factories and warehouses, where wireless links are often unreliable, and whether robot makers beyond Microsoft adopt the routing layer.