On 25 August 2026, IZON listed an exclusive tour of Kinisi Robotics’ Bristol headquarters with CEO Brennand Pierce. Bear Robotics reshared that tour the same day. Two days later, on 27 August, Korean industrial daily TheGuru (reporter Kim Yeji) cited IZON and said the KR1 wheeled humanoid is now running a crankshaft-parts assembly pilot at that same HQ, for an unnamed global automaker.
The three jobs TheGuru printed are pick and place on crankshaft parts, insert the bearing shells, then align and press them home. That is a different cell from the tote-and-unload auto talk Kinisi was giving last year. It is still a headquarters pilot. No named plant. No published cycle time. No first-pass yield. Call it an audition, not a hire.
Bearing shells are a force job. The spec page still says TBC.
A bearing shell is not a grocery tote. The insert has to seat in a recess at a consistent force. A shell that is a few tenths off will still look fine to a camera and still wreck a journal. Traditional cells put a force-torque sensor in the wrist for that reason: the resistance change is the inspection. I have not seen a public KR1 wrist-sensor sheet, and I will not invent one.
A headquarters cell also hides the hard part of automotive work. Lighting is known. The fixture is known. The SKU is probably one crank and one shell family. A plant floor adds batch-to-batch journal variation, oil and swarf on the bore, and a cycle the rest of the line will not wait for. IP65, which Bear prints on the assembly pitch, is a dust and splash rating. It is not a force spec and it is not a cycle-time spec.
Bear’s own Kinisi page, live as of this writing, is the document I will use. Form factor: wheeled humanoid mobile manipulator. Jobs: pick, place, sort, move. Payload: to be confirmed at launch. Reach: to be confirmed at launch. Mobility: autonomous omnidirectional base, also marked to be confirmed at launch. That is the vendor talking. Treat every third-party payload or arm-count table you see this week as someone else’s spreadsheet until Bear fills those three rows.
TheGuru, citing IZON, printed a top speed of 2.4 m/s. Bear’s Kinisi page says wheels that outpace legs, up to 14.4 km/h. That is 4.0 m/s. Those are not the same number. I am not going to average them. One is a reporter’s figure from a tour. One is the vendor page. Neither one is a joint-torque curve, a continuous payload at reach, or a press-fit force window.
Wheels are the right factory argument. They are not a press-fit.
Pierce has been making the same mechanical case since 2025. In The Robot Report’s “5 questions” interview he asked why a warehouse needs 12 or 14 leg motors when two cheap wheels will do on a flat floor. That argument is still the best thing Kinisi has. Most auto assembly floors are flat. Balance current is wasted current. Battery and compute should sit in the arms that actually touch the part.
The crankshaft claim tests the other half. Bear now sells Kinisi as learn-by-demonstration: show the task once, drop the robot in, skip months of integration. TheGuru said operators can demonstrate over remote or VR. Fine for a tote of known mass on a known shelf. A journal bearing is a contact problem. If the policy is mostly vision, seating is below the camera. If a data-collection glove captured force, publish the plot. Nobody published that plot.
Bear also says Kinisi is already sorting glass in live deployments, by shape, size, and material, and that the assembly pitch is IP65 for dust and splash. Glass sort is repetitive, vision-heavy, and forgiving of a missed grasp. You drop a bottle, you pick another bottle. You cock a bearing shell, you scrap an engine. Do not let a recycling cell stand in for a powertrain line.
The same page promises fleet learning: what one robot learns on a Tuesday, the fleet knows by Wednesday. That is a cloud slogan until someone publishes a transfer result on a second crankshaft SKU, in a second fixture, with a second batch of shells. I will believe the tote. I will wait on the insert.
Fifteen thousand processes is not a purchase order
Pierce, via TheGuru citing IZON, said the automaker had identified more than 15,000 similar automatable processes across its global sites, and called that a roughly US$2 billion opportunity from a single customer. That is TAM talk. No customer name. No plant. No unit count. No cycle. No yield. TheGuru also printed a build cost of £30,000 to £40,000 per robot and a robots-as-a-service figure around US$2,500 a month. Those numbers are not on Bear’s public Kinisi page. Treat them as reported interview figures until a quote sheet exists.
The corporate stack around the pilot is real enough. Bear announced a definitive agreement to buy Kinisi on 22 June 2026, opening day of Automate in Chicago. Korean financial coverage put the equity at about ₩13.2 billion. The Robot Report said the KR1, the Bristol team, and the manipulation models (a vision-language-action model plus a robot foundation model) fold into Bear’s existing fleet stack. LG already controls Bear. None of that seats a bearing at automotive tolerance.
A headquarters demo for an unnamed OEM is how every mobile manipulator gets its first auto slide. Automotive process engineers already have the test: repeatability across batch variation, at the cycle the line is already running, with a documented stop and a documented force window. Until those numbers exist, KR1 has a better factory story than a sprinter into a pad, and a thinner datasheet than a cobot that has been pressing shells for a decade.
FIRGELLI does not make Kinisi’s wheel modules or its unpublished arm joints, and this is not a claim that we do. Linear actuators belong on the fixture around a press-fit cell: the locate, the hatch, the lift that presents the crank. If you are comparing KR1 to Atlas, Optimus, Digit, or a Cruzr-class wheeled unit, hold it to the same bar. Peak and continuous joint torque. Rated payload at reach. Continuous force at the insert. A named plant. Bear left payload and reach blank on purpose. Fill those rows before you believe the US$2 billion sentence.