Somebody Is Driving: Teleoperation Is the Quiet Term in Autonomy Economics

Deployed robotic systems — delivery devices on pavements, vehicles on roads, machines in warehouses — are typically autonomous in the sense that they handle the ordinary case without instruction, and are supported by remote human operators for the rest. The support takes several forms, from direct control at low speed to a lighter kind of assistance where the machine has stopped, has proposed what it intends to do, and needs a person to approve it. The distinction matters technically and legally, but the commercial arithmetic is driven by something simpler: how much human attention each machine consumes per hour of operation. That ratio is the business. A fleet where one person can oversee a large number of machines is a technology company with attractive margins. A fleet where the ratio approaches one to a few is a staffing company with a hardware cost, and no amount of capability elsewhere compensates. The ratio is also where the improvement curve actually shows up: progress in these systems is often better described as a falling intervention rate than as new capability, and it is the metric an operator watches even when it is not the one that gets announced. Interventions are not distributed evenly, which is what makes staffing genuinely hard. They cluster — on weather, on unusual events, on roadworks, on the seasonal reorganisation of a warehouse — so the fleet needs operators available for the peak rather than the mean, and the peak is when everything needs help at once. A system sized to average demand appears adequate for months and fails on the first bad day, in public, with every unit stopped simultaneously. The operator's own working conditions are an underrated engineering constraint. Remote supervision is monitoring work punctuated by sudden demands for precise judgement, which is the pattern human attention handles worst, and the interface must present enough situational context for a good decision within seconds, over a network link with real latency. Serious deployments treat this as a design problem with the same weight as perception: how many machines one person can watch before the quality of intervention decays, how handover works, and what the machine does when the link degrades — because the correct behaviour on losing contact is not "continue", and getting it wrong is how a good week becomes a public incident.