MS Enhances Battery Life and Performance by Offloading Physical AI Robotics Inference

Microsoft Research demonstrated the impact of offloading inference operations from onboard GPUs to the edge or cloud in physical AI robotics. Previous approaches relied on GPUs inside the robot, but as models grew larger, issues with power consumption, weight, and cost escalated, limiting scalability.

Performance and Battery Improvement Effects

In evaluations of mobile manipulation tasks, distributing inference externally yielded the following benefits:

• Performance Improvement: Small onboard GPUs were up to 383% slower in mapping and planning speeds compared to A100s, and navigation obstacle detection accuracy dropped by 30%. VLA models also saw a 50% decrease in accuracy.
• Battery Life Extension: Large onboard GPUs like the Jetson Thor increased battery consumption by up to 160%. Replacing them with a Raspberry Pi-5 and offloading inference significantly improves battery life.

Physical AI Toolchain Updates

Microsoft added inference offloading capabilities to the Physical AI Toolchain. This tool automatically deploys and distributes AI workloads between robots, edge, and cloud via Kubernetes-based containerization and orchestration. It is compatible with ROS2 and LeRobot and provides real-world robot example projects such as SO-101 and UR10e.