Cloud-Edge-Device Collaboration

Splitting a robot's compute across the cloud, an edge server, and the robot itself, with each layer handling a different job.

This is an approach to splitting compute across layers: the cloud handles training and any large-model inference that needs a lot of GPU memory, an edge server (a local machine in the same building or facility) handles low-latency vision and planning, and the robot's own onboard controller runs only the control loops that absolutely must be real time. This split exists because a large embodied-AI model's compute needs conflict with a robot body's constraints on power, heat, and cost: running everything onboard is often too slow, but running everything in the cloud means a network hiccup can cause a loss of control. A common pattern is a remote policy server for inference, with locomotion control on the robot itself as a fallback — directly related to on-device deployment, inference latency, and asynchronous inference.

Exampleopenpi, the open-source implementation of π0, provides both a policy server and a client: the policy runs on a GPU-equipped machine, while the robot side only sends observations and receives actions.

Also called

cloud-edge collaboration

Related

On-Device / Edge Deployment · Policy Server (Remote Inference) · Inference Latency · Asynchronous Inference · On-device Model · Onboard Compute Platform

Sources

Edge computing - Wikipedia
openpi (Physical Intelligence) - GitHub