Robotic hardware has become dramatically more capable and considerably cheaper. The constraint has moved: for most industrial deployments, what limits usefulness is no longer what the machine can physically do, but how well the software understands the environment and how easily a person can supervise it.
Perception is a data problem
Models generalise poorly to the specific mess of a real site. The work is less about architecture selection than about collecting representative data, measuring failure honestly, and knowing which conditions the system should refuse to operate in.
Teleoperation deserves real design
Almost every deployed system needs a human to take over occasionally, and the quality of that interface determines whether operators trust the automation. Latency budgets, clear state, unambiguous stops and a display that does not require the operator to reconstruct context are the whole job.
Fleets change the questions
- Task allocation and scheduling matter more than the capability of any single unit.
- Over-the-air updates need staged rollout, because a bad update across a fleet is an outage with wheels.
- Telemetry has to be good enough to diagnose a fault without visiting the machine.
Simulation ties it together. A digital twin good enough to run regression tests against turns robotics work into something resembling normal software engineering — which is the point.