Microsoft has released open-source software for running a robot's AI model on a separate GPU server instead of relying solely on a GPU aboard the machine. Announced on September 23 as part of its Physical AI Toolchain, the offload component includes deployment code and examples for two robot arms. It could give developers access to larger models without adding as much onboard compute, provided the connection to the server is fast and reliable enough for the task.
Illustrative image of an unrelated industrial robot arm, not Microsoft's test system. Image: Lexington Medical, Inc. / Wikimedia Commons, CC0 1.0.
Robot control depends on repeated model inference: observations go in, and the model returns information used to choose an action. A March Microsoft Research measurement study found that its full mobile-manipulation workload would not fit on smaller onboard GPUs. Larger onboard GPUs consumed more battery power. The same study found a counterweight to remote computing: network delay reduced task accuracy, while straightforward cloud offloading demanded too much bandwidth.
In Microsoft's implementation, a robot sends sensor data to a control application in a remote Kubernetes cluster. Developers opt a workload into offloading with a label and configuration file. A controller then creates a GPU server pod, and the runtime sends selected model calls from the control application to that pod and returns the results. The robot's sensor and actuator connection to the cluster must be supplied by the operator; the tool does not create it.
The repository includes reference examples for an SO-101 arm and a UR10e arm. Microsoft's announcement also shows the Rho model's inference offloaded to a Jetson Thor GPU while it controls a Mobile Aloha robot in two demonstrations. The public code provides a route for developers to try the arrangement, but the repository describes the real-hardware examples as reference material that must be adapted to each setup. Its README assumes an isolated GPU-server network and warns that the design does not follow all conventional enterprise security practices.
