We build task-specific robot cells for small and mid-size manufacturers — machines that learn a job by demonstration instead of waiting six months on a custom integration. You show the robot the task. It runs the task.
Fixed tooling amortizes over hundreds of thousands of identical cycles. It assumes one part, one orientation, one presentation, held to a fixture built for that part alone. Under those conditions it is unbeatable — and those conditions describe a minority of American manufacturing.
The rest of the floor looks different: forty SKUs, seasonal runs, a changeover on Thursday, a bottle that gets a new label in the spring. Every one of those variations is a re-fixturing charge against a machine that only pays for itself by never changing. So the quote comes back at six figures, the payback lands past sixty months, and the plant is told — politely — that automation isn't for them.
We think that conclusion was correct about the tooling and wrong about the plant. The constraint was never the factory. It was that the machine had to be told the task in coordinates instead of being shown it.
A conventional cell encodes a task as a list of poses. Ours encodes it as a policy — a learned mapping from what the cameras see to what the joints should do next.
An operator drives the cell through the real task with a leader arm at 50 Hz. The robot logs its own wrist and overhead camera streams alongside joint states and commanded targets. Every episode is a labelled example of the job done correctly, on your parts, in your light.
Because the policy is conditioned on vision, tolerance to part placement is bought with demonstrations rather than with fixtures. A bottle two centimetres off-centre isn't a fault condition. It's a case the policy has seen before.
The trained policy runs on hardware in your building — off-robot, but on-site. No cloud round trip, no per-cycle inference bill, no line stoppage because the ISP dropped. The control link is budgeted in milliseconds, and we publish that budget.
What this does not mean. A learned policy is not a general-purpose worker. It is very good at one task it has been shown many times, and it degrades outside that distribution. We scope cells to a single well-defined motion and say so plainly, because the failure mode of overselling this technology is a stopped line.
We time the task on your floor with a stopwatch and a tape measure. Reach, deck height, presentation, current cycle, current scrap. Measured, not estimated.
Work envelope, payload budget and duty cycle are written down before anything is quoted — including an explicit list of what the cell will not do.
Teleoperated episodes recorded on the live line with production parts. Enough repetitions to cover the presentations the task actually produces, not the ones it produces on a good day.
Policy training with held-out episodes reserved for evaluation. A policy that has never been scored against data it didn't train on has not been evaluated.
Supervised running at rate. We report success rate, mean and worst-case cycle time, and the observed failure modes as measured — not as targets.
Your operators are trained to record new demonstrations themselves. Products change. The policy should be re-teachable by the people on the floor, not by us on a service call.
King Robotics shares ownership with a working distillery in Centralia, Washington, producing at roughly 800 gallons a month. The bottling line is real, the labour on it is ours, and the first cell we built had to survive it. We are our own first customer and our own first reference — which is the only honest position for a company this young to take.
The task: lift empty bottles from an untaped twelve-count case, present them to the labelling conveyor, recover them after label, and return them to the case.
| Manipulator | Bimanual, 19 DoF, mobile base |
|---|---|
| Payload | 1.5 kg per arm · 3.0 kg bimanual |
| Reach | 550 mm · column travel 30 mm/s |
| Perception | 4 × RGB — wrist L, wrist R, overhead, front |
| Grasp strategy | Neck grip, current-estimatedno force feedback |
| Control link budget | 150 ms LAN · 300 ms WAN · motion halts past budget |
| Inference | On-premise, off-robot |
| Thermal limit | 60 °C servo cutoff |
| Status | In build — first policy Q4 2026on schedule |
Case handling stays human. A loaded twelve-count case weighs 5–8 kg against a 3 kg bimanual payload. Palletising is not a roadmap item on this cell; it is physically out of reach, and we publish that before a quote rather than after a deposit. If a limit like this kills your application, we would rather you learn it from us in the first meeting.
No proposal leaves here based on a description of the task over the phone. If we haven't timed it, we haven't priced it.
Inference runs on-site. Demonstration footage of your production line is your footage — it is not published, not pooled into a marketplace, and not used to train anything we sell to anyone else.
Payload, reach, latency budget, and the tasks a cell will refuse are written into the scope document in the same type size as the capabilities.
A cell your team cannot re-teach is a dependency, not an asset. Handover isn't complete until someone on your floor has recorded a policy update without us in the room.
That single repeated motion is the right size for a first cell. Send it over and we'll tell you honestly whether it's a fit — including when it isn't.
Centralia, Washington · Serving the I-5 corridor
What makes a survey fast