Remote GPUs help robots, but the network controls their reflexes

Microsoft compared mobile robotic manipulation across onboard, edge and cloud GPUs. A smaller Jetson Orin with 32 GB of memory could not fit the full stack. On weaker hardware, mapping and planning were up to 383% slower than on an A100, timely obstacle detection fell by 30%, and VLA model accuracy dropped by as much as 50%.

Remote GPUs relieve both the battery and the model

According to the study, a more powerful onboard Jetson Thor increased battery drain by up to 160%, costing several hours of operation. Moving inference to a more capable external GPU enabled larger models and improved task success. Researchers tested SO-101, TurtleBot 4 and Stretch 3 robots across perception, planning, navigation and manipulation.

Compute becomes part of the robotics infrastructure

For fleet operators, this changes the economics of the machine. A robot does not have to carry the most expensive GPU or its power draw, while edge compute can be shared. The cost moves into networking, capacity planning and task prioritization.

That matters most outside tightly controlled factories. A robot in an office or home must continuously react to people and obstacles, so remote performance is valuable only if it arrives before the physical action.

Network delay collects its fee in accuracy

Offloading is not free acceleration. Just tens of milliseconds of additional latency reduced manipulation accuracy by more than 10%. Continuous video transmission can saturate the network, while compression cut manipulation and semantic mapping accuracy by nearly 20% in the study.

A shared GPU also creates a queue. When several robots call it at once, inference and network latency rise, so critical tasks need admission control or guaranteed priority.

Fleet operations will choose between edge and cloud

The next meaningful signal will be the behavior of an entire fleet rather than one robot in a lab test. Watch guaranteed response times under load, transmitted video volume, battery life and the quality of safe local fallback during an outage. Those metrics will show whether offloading belongs in open environments.

Lilith's verdict

A robot can carry a lighter body, but its reflexes then live in the network. When the connection stutters, the expensive cloud GPU merely watches from afar as the machine misses the obstacle.