Microsoft Research's robot GPU platform study

Microsoft Research published a measurement study comparing onboard, edge, and cloud GPU platforms for mobile robotic manipulation, with results showing that smaller onboard GPUs are insufficient for running full AI workloads. Lighter onboard GPUs slowed mapping and planning by up to 383% compared to an A100, and navigation showed a 30% drop in timely obstacle detection. Some smaller GPUs lacked enough memory to run the full mobile manipulation stack at all. Larger onboard GPUs can handle the workload but drain robot batteries several hours faster. Offloading inference to the cloud improved response time, accuracy, battery lifetime, and cost, but cloud offloading introduces network latency that degrades task accuracy, and bandwidth requirements make naive cloud offloading impractical. The study also found that sharing compute across robot fleets presents measurable opportunities but also pitfalls.