Offloaded inference for real-world physical AI robotics
Microsoft says robots can perform better when the heaviest AI inference runs off the machine.
Its study found onboard GPUs can cut battery life, add weight and cost, and limit which robotics models can run. In mobile manipulation tests, offloading to edge or cloud GPUs improved response time, accuracy, and task success, including in mapping, navigation, and handover work. Microsoft is adding Kubernetes-based offloading support to its Physical AI Toolchain so developers can split robotics workloads across the robot, edge hardware, and cloud. Microsoft Research AI's note
Its study found onboard GPUs can cut battery life, add weight and cost, and limit which robotics models can run. In mobile manipulation tests, offloading to edge or cloud GPUs improved response time, accuracy, and task success, including in mapping, navigation, and handover work. Microsoft is adding Kubernetes-based offloading support to its Physical AI Toolchain so developers can split robotics workloads across the robot, edge hardware, and cloud. Microsoft Research AI's note
score 5