General-Purpose Robotics

The distance between a system that can do a task and a system that will do it ten thousand times without help is where robotics actually lives.

Robotic manipulation has genuinely improved. Systems now perform task families that were out of reach three years ago, and they generalise across object instances in ways that earlier approaches did not. This paper does not dispute that. It asks a narrower question: what happens to those capabilities under sustained operation in uncontrolled conditions, and what does the answer imply for deployment.

The intervention gap

We collected intervention logs from twenty-three operational deployments and compared them with the same or comparable systems under demonstration conditions. Mean time between human interventions falls by between one and two orders of magnitude. A system demonstrating a thousand consecutive successful grasps in a laboratory may require intervention every few dozen attempts on a production floor.

Where the failures come from

The failure distribution is the most useful part of this dataset. Across deployments, the majority of interventions trace to perception rather than control: an object not recognised under unusual lighting, a partially occluded target, a reflective surface, an unexpected item in the workspace. Control failures — a grasp that slips, a trajectory that collides — are a minority.

This matters because it locates the problem. Perception under distribution shift is the sort of problem that responds to data and to scale; several deployments in our sample show meaningful improvement over eighteen months on exactly these categories. The improvement is real. It is also slower than the rate at which task generality has expanded, which means each new capability arrives at a reliability level below what deployment requires.

The economics as they stand

For every deployment where we could construct a cost per successful task, single-purpose fixtured automation remained cheaper than general-purpose manipulation. The gap narrows with task variety — general-purpose systems win where the task mix changes frequently enough that re-fixturing dominates — but that regime is narrower in practice than the case for general-purpose robotics usually assumes.