Microsoft Research published a measurement study on 23 September arguing that the graphics processor bolted onto a robot is usually the wrong place to run its models, and that the work should be sent to a nearby server or the cloud instead.
The numbers are Microsoft’s own, taken across several robots (SO-101, UR10e, Mobile Aloha and Stretch-3) and several compute options from a Raspberry Pi 5 up to an NVIDIA A100. Even on onboard chips with enough memory, mapping and planning ran up to 383% slower than on an A100. Navigation lost 30% of its timely obstacle detections on lighter chips. Vision-language-action models did not slow much, but slowed enough to halve their accuracy. And the bigger onboard chips are the ones that drain the battery: Microsoft reports a Jetson Thor costing a Stretch-3 up to 160% of its battery life, or a few hours of running time.
The catch is stated plainly in the study: offloading trades one constraint for another, and now depends on network latency, bandwidth and whether a spare server GPU is actually available. A robot that stops working when the Wi-Fi drops is a different engineering problem, not a solved one.