Cognex buys RealSense, Agility eyes wheels

Today: Cognex buys a depth sensing platform many robot stacks rely on, Agility questions its form factor, Tesla’s humanoid line meets the limits of robot hands, and two papers stretch demonstrations further.

Cognex to buy RealSense for about $500 million

Cognex announced a definitive agreement on 23 September to acquire RealSense, the depth sensing company Intel founded in 2014 and spun out in 2025, for about $500 million in cash from its balance sheet, with a staff retention program folded into the terms and closing expected in the fourth quarter of 2026. RealSense will spin its facial authentication line into a separate company before the deal closes.

Cognex sizes the robotic perception market at about $600 million today, growing more than 25 percent a year to roughly $1.6 billion by 2030, and expects RealSense to post $80 million to $90 million of 2026 revenue, up more than 50 percent year on year. Chief executive Matt Moschner describes a full stack running from industrial ID and 2D and 3D measurement to depth perception. RealSense chief executive Nadav Orbach lists perception guided arms, mobile robots, quadrupeds and humanoids as its markets.

For developers the question is continuity, not strategy. RealSense depth cameras and their ROS drivers are a common starting point for perception on mobile robots and manipulators, and a machine vision incumbent running that platform can mean better calibration and factory tooling, or a roadmap steered toward enterprise accounts. Watch whether the SDKs keep moving, and remember that camera and compute choices are usually made together, which is why Jetson Orin Nano versus Orin NX is the first comparison teams run.

Source: Quality Magazine.

Agility is exploring wheeled robots

A wheeled robot appeared near the end of Agility Robotics’ Digit 5 launch video, and The Robot Report asked chief executive Jonathan Hurst why. His answer: “We’re exploring a range of robot designs and form factors, including wheels, to address different operating environments and customer needs,” adding that Agility wants “to bring robotics to more environments sooner.” He would not say whether a wheeled product would be a Digit variant or a new robot, and there is no timeline.

The trade is familiar. Wheels are cheaper and more stable on flat, predictable floors, while legs stay the answer where stairs, obstacles and human shaped layouts decide the task. The notable part is who is saying it: a company built on bipedal locomotion is keeping its options open. The commercial pressure is real, with a merger with Churchill Capital Corp. XI valuing Agility at $2.5 billion and expected to raise more than $620 million in gross proceeds, and an S-4 filing reporting more than $300 million of multi year Digit v5 orders from one customer.

Base choice is not cosmetic for developers. A wheeled base brings differential drive constraints, simpler state estimation and an easier safety case, and it changes what manipulation has to plan around. Treat the direction as company intent, not product.

Source: The Robot Report.

Tesla’s Optimus line meets its hands

Reporting by The Information, summarised by Ars Technica, describes Optimus V3 as hard to build. Fremont stopped making the Model S and Model X in May 2026 and moved those workers to Optimus, but assembly equipment struggles to line up components, the line cannot run as fast as planned, and each hand and forearm holds more than 100 small parts that people still assemble by hand. Production has reportedly reached hundreds of robots per week against a target above 1,000 per week by the end of 2026.

Two details matter for manipulation work. Touch sensors were unreliable enough that Tesla built a glove like sensor layer that can be replaced without swapping the whole hand, a sign finger level sensing is still the least mature part of a humanoid. And the data story is a labour story: workers were asked to wear suits recording their movements for imitation learning, some objected to training their replacements, and the company moved to dedicated collection teams.

The bottleneck has moved from the model to the hand and the dataset, and both iterate at the speed of hardware. See why VLA models matter for the scale argument behind the data push.

Source: Ars Technica.

Amazon adds robot manufacturing in Indiana

Amazon will invest more than $100 million in a 585,000 square foot advanced manufacturing facility in Greenwood, Indiana, adding 300 manufacturing and engineering roles and starting production by 2028. The plant will build products that power Amazon’s fulfilment and robotics network in North America, combining fabrication, robotic welding, automated powder coating and assembly under one roof, and follows an August announcement of robotics manufacturing investment in Texas.

Amazon says it has manufactured more than one million robots in the United States, that they work across more than 300 facilities globally, and that they assist employees with 75 percent of customer orders. Roboticists rarely build the machines they program, but upstream capacity sets their constraints, and more domestic fabrication shortens the loop between fleet data and the factory.

Source: The Robot Report.

Policies without training, from stored demonstrations

A Stanford group (Maximilian Adang, Timothy Chen, Lars Osterberg, Aiden Swann and Mac Schwager) published Training-free Behavior Cloning on 24 September. Their Behavior Predictive Control method skips end to end training: it predicts the next actions by blending stored observation and action data that best reconstructs the robot’s recent history, using an action aware retrieval metric, a Hankel based action continuation prior and a closed form one step residual correction.

The authors report that the method is competitive with learned policies such as pi0.5 and better in some cases, that fitting drops from hours to seconds on consumer GPUs, and that closed loop control runs above 75 Hz on a Jetson Orin Nano. Keeping demonstrations at deployment makes each prediction traceable and lets you change behaviour by editing the demonstration bank instead of retraining. The usual caveat applies: this is the authors’ own evaluation, with no independent replication. The trade is memory for a policy you can inspect and blame, the opposite of the scaling route in why VLA models matter.

Source: arXiv.

Coding agents write robot programs from one demonstration

RAPID, from Yuyao Liu, Jiayuan Mao, David Hsu, Leslie Pack Kaelbling and Tomás Lozano-Pérez, takes a single visual human demonstration and infers three things from it: a testable task specification, the action primitives the robot can run, and an interactive environment where candidate programs can be executed and verified. It then loops on generating, testing and refining code.

The representation choice is what makes the programs reusable. RAPID stores an object centric relational program rather than a motion trace, expressing primitives as trajectory optimisation programs and composing them through relational constraints, so scene geometry is resolved at run time. The authors report simulation results on eight contact rich nonprehensile manipulation tasks plus LIBERO-Pro, and deployment on a real Franka arm across all eight. The numbers are the authors’ own. Programs give you a diff, a stack trace and a place to put a check before the arm moves.

Source: arXiv.

Quick Bites

• CNH Industrial’s Luca Ferrari frames the shift in agriculture as agentic systems that run a whole operation instead of assisting one task, with labour availability in planting, spraying and harvesting as the pressure point. The Robot Report.
• IEEE RAS traces how China’s Five Year Plans moved from Made in China 2025 to embodied AI and humanoids, driven by manufacturing depth and an ageing workforce. Robohub.
• A post-training recipe for hardware drift in VLA policies compresses observation, proprioception and action history into a context token that modulates the policy through adaptive layer norm. The authors report recovering over 80 percent of base policy performance under actuation bias and joint encoder offsets. arXiv.
• IEEE Spectrum profiles Barbara Mazzolai on sustainability robotics and treating the environment as more than design inspiration. IEEE Spectrum.