Five things a shop would otherwise buy from five vendors, integrated and qualified at HQ, and delivered as one unit that arrives running.
A Relling cell ships as one unit: arm, sensing, tooling, software, safety, qualified at HQ. The job was the same shape, four stations around one cobot, owned end to end.
Relling rebuilds American manufacturing expertise by making advanced automation practical for small and mid-sized shops. Everything in this chapter exists to say that in a way a plant manager believes: the marks, the site, the funnel that carries a lead, and the films that demystify the machine.
Marks, tokens, type, and the jacket, one system that holds on a product page, a brochure, a film, and a body on the floor.
Signal#d99f2b
Ink#1a1a1a
Canvas#f9fafb
Navy#2b3a73
I designed and built rellingsystems.com from scratch. One job: take machine vision, teleoperation and turnkey cells, and hand them to a shop owner who does not have a robotics team on staff.
Every page opens on the job to be done and earns its technical detail later. 19 capability pages and 11 industry pages exist so a search for a real shop-floor problem lands on a page about that problem, not on a home page.
Made in America. Built for America’s industrial base.
The flag above is not an image. An 8×8 Bayer matrix picks one of five glyph states per cell and a travelling sine moves the cloth; every dense cell is the Relling mark from its own vector paths. Point the same engine at a photograph and you get a plate. Drag the handle to see the input.
The workcell page leads with the machine itself, running. It boots before the page images do, because on that page the hardware is the headline. What you are watching is the real page, recorded.
One arm picks off the conveyor and hands to the other mid-air; the second packs the box for the pallet. Inverse kinematics against the real joint limits rather than a baked animation, which means the handoff has to actually rendezvous.
On the careers page it hunt-and-pecks its own application into a 1984 Macintosh, hesitates over submit, and gets a bomb dialog. Then it starts again. Same engine, same rigged arm as the two scenes above.
Skipped outright on phones and data saver, warmed only when the browser is idle, booted by an observer while still a screen away. Three.js is self-hosted: the CDN copy kept dying on phone ad blockers.
Each one pre-steps about two seconds before its first frame is shown, so it fades in on a settled pose. The camera reframes by aspect: wide is the whole bench, narrow crops to the handoff.
Meta, Google, and Microsoft plus the live site feed PostHog. Attio takes that, and conversion data goes back to the advertisers, a self-healing loop, not a sink.
Keyword SEO for the queries a shop still types. AEO so an agent cites the live site instead of inventing the product.
Pages structured so those terms resolve to a real cell, not a slogan.
Clear claims on the public pages. The site is the source, not a campaign beside it.
A cell gets bought by a room, not a person. The brochures, renders and leave-behinds exist so the one champion inside a plant can carry the argument to everyone who has to say yes.
The spec sheet a plant manager can hand to procurement and to the engineer who will actually run the cell, in the same visual language as the site.
High-fidelity stills of a cell that does not exist on a floor yet, so a buyer can see the thing before there is anything to photograph.
Cut in Premiere and After Effects when a still cannot show a cycle. One of them is playing at the top of this page.
Digital trust has to be backed by physical execution. This chapter is the part you can put your hands on: cells assembled and qualified before they ship, and the tooling modelled and printed to make a specific part move.
I designed, built, and assembled turnkey cells across a variety of robot vendors, cobots, KUKA, FANUC, and other arms, for multi-million-dollar deployments. Not one SKU. Not a render.
Arm, sensing, tooling, software and safety, qualified at HQ before any of it lands on a customer’s floor. The page on the right is the public version of the same object; the rest of this station is what it takes to make one.
Years of taking things apart and building them back is what made this end of the work possible. A cell has to boot, see, think, and stay inside a customer’s walls. I built that part too.
Spec’d and assembled the compute units, including NVIDIA Jetson modules sized to run perception and video inference at the cell rather than in a datacentre. Bench-built, burned in, and racked as part of the cell instead of arriving as somebody else’s black box.
Ran the harnessing and the routing: arm, sensing, tooling, safety and compute wired into one unit that a shop can power on. The parts of a deployment nobody photographs and everybody feels when they are done badly.
Live video capture and training run on site, on the customer’s own hardware, behind their boundary. That is a hardware decision before it is a software one: enough compute at the edge that nothing has to be shipped out to be useful.
Defense and controlled work cannot send footage of a line to a cloud. Keeping capture, storage and inference inside the facility is what lets a buyer with ITAR obligations get past the first question and actually evaluate the cell.
What the parts actually are belongs to the customer, so the drawing below stands in for them. The path is the part I can show: modeled in CAD, fabricated, materials-tested, and then bolted onto cells that run on real deployment floors, on cobots, KUKA and FANUC.
Training a vision-language-action model needs clean data and an interface an operator can actually feel through. This chapter is the software side of the arm: the controls an operator works through, built to capture the contact-rich data the models learn from.
I owned this end to end: the leader arms an operator actually holds, the stands they sit on, the software that drives them, and the recordings that come out the other side. The hardware and the software were built together, because an arm that feels wrong in the hand is usually a software problem, and the other way round.
Small leader arms assembled from an open kit: servos, linkages, the handle an operator grips. Cheap enough to build several, which matters when the point is collecting hours of data rather than owning one perfect rig.
Bimanual stands so one operator drives two arms at once. Most of the work worth automating takes two hands, and a rig that only captures one teaches the model half a task.
The driver and control loop that turn a stand of servos into something an operator can feel through: low latency, live scene state, and precision tight enough that contact-rich work stays legible instead of turning into a fight.
Sessions record to MCAP logs, the format the policy training reads. Everything above exists so that file is clean: if the capture is wrong, the hardware and the interface were wrong first.
A startup is not only its product. This chapter is the bird’s-eye view: keeping the funnel, the floor and the team pointed at one bet.
GTM, hardware, and culture pointed at the same bet, or the other eleven stops are a list of side projects.
Ads into Attio. The lead and the cell in one record.
Cells that ship tested. Multi-vendor, multi-million-dollar deployments.
Generalists with depth. The people who can hold more than one lane.

I own the pipeline from the printed tool on the wrist to the teleop that trains the model to the site and ads that bring the buyer in. Building a robotics company from zero needed a generalist, so I became one: brand, growth, hardware across vendors, software, the room, and the hire.