Does a Robot Need Its AI Brain Onboard?

1 The prevailing assumption

A Microsoft Research blog post published September 23, 2026 describes the common approach to physical AI as provisioning a GPU onboard the robot so inference stays local. The authors challenge that: GPUs consume significant power, cut battery life, add cost and weight, and can limit running the latest generation of models. Robotics combines a power source, mechanical construction, a control system, and software; onboard inference loads two of those four.

2 What offloading buys

The study focused on mobile robotic manipulation, on a canonical task such as checking the kitchen for rubbish and putting it in the trash. Reported results are specific: some smaller GPUs could not accommodate the mobile manipulation stack at all; on GPUs with sufficient memory, mapping and planning slowed by up to 383% versus an A100; navigation showed a 30% drop in timely obstacle detection. Vision-language-action models slowed less dramatically, but enough to drop their accuracies by 50%.

3 Battery is the second axis

Replacing an onboard GPU with a Raspberry Pi-5 board and shipping data to an offloaded GPU increased battery lifetime, the post reports. Larger onboard GPUs, such as Jetson Thor, drained robot batteries by up to 160%, a few hours, with the cited numbers for the Stretch-3 robot. That is a reported measurement from one study of representative hardware.

4 The cost side of the ledger

Offloading is not free. The authors describe it as a complex tradeoff involving performance, network latency and bandwidth, and available GPU resources, each a dependency the robot did not have with the GPU inside its chassis.

5 When offloading fails

The post presents no decision rule, so failure modes must be read off its tradeoffs. Offloading fails when latency sits on the safety path, when bandwidth cannot carry the sensor volume a task needs, when the site has no reliable network, or when shared GPU capacity cannot be guaranteed at peak. A robot keeping a small local model for reflex behavior and sending heavy perception outward is a different architecture, not a compromise.

6 The tooling being offered

The authors describe a toolset using Kubernetes as a uniform abstraction to distribute robotic AI across robot compute, edge, and cloud, with integration for simulators, LeRobot, and ROS2. Microsoft says it is adding this capability to its Physical AI Toolchain, an open-source framework integrating Azure services with NVIDIA's physical AI stack, demonstrated with Microsoft's Rho model offloaded to a Jetson Thor controlling a Mobile Aloha robot. That is a company claim about tooling, not a measurement result.

7 Bottom line

Offloading inference improved task performance and battery life in one reported measurement study of mobile robotic manipulation.

The price is dependence on latency, bandwidth, and available GPU capacity, all of which must hold for the robot to act.

The decision turns on whether the network sits on the safety path; if it does, the compute stays in the chassis.