Industry Brief, September 23, 2026: A vision deal, Epson's first cobot, humanoid school

By MillBrief Editorial · September 23, 2026

Four stories today, and each one is really a question about who you end up depending on. A machine-vision company is buying a depth-camera maker whose products sit inside robot arms, mobile robots and humanoids. A SCARA specialist has released its first collaborative arm. A humanoid maker has opened a permanent training centre inside its parent company’s car plant. And a contributed column asks a question buyers rarely ask at purchase time: what happens to the robot when it is retired? As always, the writing is ours, each item links to the original outlet, and vendor numbers are labelled as vendor numbers.

Cognex agrees to buy RealSense, the depth-camera maker Intel spun out

Cognex, the industrial machine-vision company, has signed a definitive agreement to acquire RealSense for approximately $500 million, funded entirely from existing cash and investments. RealSense was founded inside Intel in 2014 and spun out in 2025; its depth cameras are used in perception-guided robot arms, mobile robots, quadrupeds and humanoids. On top of the purchase price, Cognex expects to fund a three-year cash retention program for RealSense staff of $56.5 million at target and to grant restricted stock units worth about $50 million. Closing is expected in the fourth quarter of 2026, subject to the usual conditions. Before closing, RealSense will spin its facial authentication product line out as a separate company.

Figures from the release: RealSense expects 2026 revenue of $80 million to $90 million, more than 50% growth on the prior year, and Cognex sizes the robotic perception market at about $600 million today, growing more than 25% a year to roughly $1.6 billion by 2030. Our arithmetic: the headline price is about 5.6 to 6.3 times RealSense’s expected 2026 revenue ($500M divided by $90M and by $80M).

Why it matters: If a cell you run, or one you are quoting, uses a RealSense camera for bin picking, pallet detection or mobile-robot navigation, its supplier is about to change owners. Acquisitions often bring changes to product lines, SDK licensing, distribution and pricing, although nothing about any such changes was announced here. Two sensible steps: ask your integrator which camera models and SDK versions your cells depend on, and before you sign a new quote, ask how long the vendor commits to supplying and supporting those parts. Parts supply belongs in the lifetime-cost line items covered in the hidden costs of automation, and it is worth adding to our integrator question list.

Source: Robotics Tomorrow, September 22, 2026 (Cognex press release).

Epson launches its first collaborative arm

Epson Robots has added the AX6, a collaborative six-axis arm, to its AX Series. Epson describes itself as the world’s number one SCARA maker and says it has sold more than 250,000 robots (its internal figure, through April 2026). The AX6 uses power- and force-limiting, which Epson says can allow operation without traditional safety barriers with a proper risk assessment. Other features Epson lists: a carbon-fibre structure, IP54 and ISO Class 5 cleanroom ratings, no-code programming alongside a Python environment that accepts custom libraries, teach buttons plus an Ethernet port and pneumatic connections at the end of the arm, a 3D simulator, and support for 100 to 240 V AC or 48 V DC power. It is sold through Epson’s distributor channel. The release gives no price, payload or reach.

Why it matters: Epson’s pitch is a single supplier across fenced industrial arms, cobots and the in-between, which appeals to shops that would rather deal with one robot vendor and one spare-parts relationship. The wording also deserves a careful read: “can allow operation without traditional safety barriers” depends on the risk assessment, and that assessment covers the whole application, including the gripper, the part and the speeds, not just the arm. Budget for it. With no payload or reach published yet, compare the AX6 against the models on our cobot price tracker once Epson publishes its specifications, and see cobot vs industrial robot for when a fenced arm is the cheaper answer.

Source: Robotics Tomorrow, September 22, 2026 (Epson press release).

Boston Dynamics opens a permanent “robot behavior factory” at Hyundai’s Georgia plant

Boston Dynamics has opened its Robotics Metaplant Application Center (RMAC) at Hyundai Motor Group Metaplant America near Savannah, Georgia. Chief product and technology officer Zack Jackowski told The Robot Report that earlier trials were time-boxed and hands-on with the engineering team. RMAC, by contrast, is permanent: real parts, fixtures and workcells plus Atlas robots and skilled pilots, producing training data. One task it names is parts sequencing: taking automotive parts out of shipping packaging and placing them, in order, into dense cubbies for the line. That work needs two-handed handling and care over which surfaces may be gripped. Training data comes from teleoperation, reinforcement learning in simulation, and Universal Manipulation Interface (UMI) devices.

The timeline, according to the company: data collection scales through 2027; RMAC moves next year into a new building about ten times larger; other industries will be explored next year; component assembly is targeted by 2030. Hyundai has said it will deploy Atlas across its plants, starting with 25,000 units at Hyundai and Kia over the next few years, and will build a U.S. factory able to make up to 30,000 robots a year by 2028. Jackowski’s most useful remark for buyers: the team is learning less about the robot than about the application, including how fixtures and packaging might change to suit humanoids and how to build support processes.

Why it matters: Even with a parent company as its first customer, Boston Dynamics is building a dedicated facility to turn a single application into reliable robot behaviour. A smaller plant will not get that treatment, so treat humanoid proposals as application-development projects and budget for changes to fixtures and packaging. The same lesson holds for conventional automation: what to automate first matters more than which robot you pick.

Our read: We expect Atlas work outside Hyundai’s own plants to remain at the pilot or exploration stage through the end of 2027. That rests on the company’s own timeline (training through 2027, other industries explored next year, component assembly by 2030). A sign we’re wrong would be a named non-Hyundai customer running Atlas in regular production, not a pilot, before 2028.

Source: The Robot Report, September 21, 2026.

Retiring a humanoid is a job for specialists

A contributed column in The Robot Report asks what happens to a humanoid at the end of its life. The author is a business development consultant who works with Re-Teck, a recycling company, so read it as an interested party’s view. Their figures: a humanoid has 10,000 to 15,000 components, 20 to 40 electric motors, 50 to 200 sensors, and 3.5 to 4 kg of neodymium magnets, which they say can exceed the amount in an electric-vehicle chassis. They list four problems. Stored data (maps, logs, camera recordings) has to be destroyed or cryptographically erased. Lithium battery packs risk thermal runaway if crushed. Reused structural parts may carry fatigue risk. And magnets have to be removed by hand before shredding, because crushing contaminates them.

Why it matters: Most of this applies to any networked robot or mobile robot on your floor, not only humanoids. Two questions are worth asking at purchase time rather than at retirement: how the vendor wipes stored data and whether it will certify that, and who takes the battery packs back, and at what cost. Neither usually appears on a quote, which is why end-of-life belongs in the lifetime-cost view used by our TCO calculator.

Source: The Robot Report, September 20, 2026 (contributed column).