Local interpretation and execution
A robot may need nearby processing to interpret observations within a task’s decision window or to operate within a defined communication condition. This does not mean every model belongs onboard. Available memory, computing capacity, power, cooling, and competing processes limit what the platform can sustain.
Measure the actual time from observation to useful output under representative load. An average inference time alone leaves out buffering, preprocessing, scheduling, and communication. A model that fits in memory can still miss the relevant decision window or deprive another component of computing resources.
Shared task coordination
A fleet service can consider assignments, destination availability, and task progress across platforms. Its view depends on reports that may be delayed or incomplete. Treat each report as information with a source and effective time rather than assuming the service has a perfectly current global state.
Central analysis and reviewed learning
Collected records can support analysis across missions and equipment. Keep device, sensor, software, map, and model versions associated with those records. A central dataset can contain differences in physical conditions that make simple aggregation misleading. Fleet wide data availability does not remove the need for valid labels and an appropriate comparison.
Shared models do not mean automatic shared learning
Sending one model to several robots is different from training jointly across them. A learned update needs its own data permissions, evaluation, version control, and release decision. Do not describe ordinary log collection or model distribution as federated learning unless the actual training arrangement supports that description.