Hey all! ECE undergrad here, exploring an inference-routing idea for robots (deciding at runtime whether a perception or higher-level AI task runs onboard, on a nearby edge server, or in the cloud, based on deadlines, network conditions, and local compute load) and trying to validate the problem before building anything.
For those running real robots: is your local/remote split for perception or AI tasks fixed, or does it ever change dynamically? Has latency or local compute contention ever actually caused a problem, what happened, and how did you fix it (changed placement, went async, added local compute, something else)? And has anyone tried offloading inference and then walked it back. What made it not worth it?
Not asking about anything safety-critical (e-stops, motor control), just perception/AI tasks. Production, lab, or abandoned-experiment stories all welcome, just mention which is which. Thanks!