Robotics Weekly Review 2026-08-29

Week In Review

The dominant story of the past week was not a new robot but a new economics. Three separate announcements — XPeng’s robotics unit raising more than $900 million, Gatik’s $200 million Series D, and Bedrock Robotics putting operator-free excavators onto paying job sites — describe an industry that has stopped raising money against demos and started raising it against contracted revenue and delivery counts. Underneath them, NVIDIA’s Jetson Orin Nano 2 quietly moves the floor: doubling edge inference throughput in the same small, low-power package is what lets a large vision-language model run on a drone or a floor robot rather than in a data center it cannot reach.

A second thread ran through the week with unusual clarity: the binding constraint on physical AI is data, not model architecture. Carbon Robotics’ new plant foundation model works because it was pre-trained on 150 million labeled plant images, a corpus assembled with a specialist annotation partner. Two days later that same partner, iMerit, was acquired by EXL — a data-services company being bought at the moment its output became strategic. And Locus Robotics, explaining why it acquired a gripper startup, made the same point from the hardware side: tactile sensing is hard precisely because simulation does not produce usable touch data. Three companies, three vantage points, one conclusion.

The week’s most concrete pattern, though, was retrofit rather than replacement. Bedrock installs a sensor-and-compute kit on a conventional excavator in a single day with no permanent modification. Carbon Robotics released an autonomy kit that converts John Deere tractors built since 2019 into driverless weeders. Gatik runs ordinary box trucks on fixed regional routes. None of these companies is building a new machine; each is removing the person from the seat of a machine that already exists and already has a business model. That is a far cheaper path to deployment than the humanoid route, and it is where the operating hours are accumulating fastest.

Public institutions moved in the same direction. The NSF funded a five-year center at UT Austin devoted to how people and robots adapt to each other — an explicit bet that the hard problems in service and assistive robotics are relational, not mechanical. The Department of Transportation doubled the FAA’s BEYOND drone program and pointed it at heavier cargo and higher altitudes. And at the far end of the time horizon, NC State researchers published a soft robot that leaps continuously under nothing but infrared light, with no motor, no battery, and no reset between jumps — a reminder that the field still has room for machines built on entirely different principles from the ones now raising nine-figure rounds.

Items

XPeng’s Robotics Unit Raises Over $900 Million at a $6.3 Billion Valuation

XPeng Motors spun its humanoid robotics work into a separate business, Dogotix, and closed a first external financing round of more than $900 million. The round was led by IDG Capital, with Gaorong Ventures participating and Alibaba and Tencent among the strategic investors. XPeng reports a pre-money valuation of $5 billion and a post-money figure above $6.3 billion, and characterizes the deal as a record for a single private financing in China’s embodied AI sector.

What separates this from the general run of humanoid funding is the manufacturing context. XPeng is an automaker, and the pitch to investors is that the unit can borrow an existing vehicle-scale supply chain, factory footprint, and autonomy stack rather than assembling those from scratch. The company says the money will go toward hardware and software development, data collection, training physical AI models, building mass-production facilities, and international expansion — in that order, which is itself informative about where the costs actually sit.

XPeng expects its next-generation IRON humanoid to enter mass production by the end of 2026. That is an aggressive date, and worth treating as a target rather than a schedule. But the underlying claim is credible in a way it would not have been two years ago: several Chinese manufacturers have now demonstrated that humanoid hardware can be built in volume, and the open question has shifted from whether the machines can be produced to whether there is work for them that justifies the unit cost.

The deal also sharpens a structural asymmetry in the sector. Chinese manufacturers held the large majority of global humanoid shipments through the first half of 2026, and a round of this size — raised domestically, from domestic strategic investors, against a domestic manufacturing base — suggests that lead is being reinforced rather than contested.

Source: The Robot Report

NVIDIA’s Jetson Orin Nano 2 Doubles Edge Inference in the Same Power Envelope

NVIDIA announced the Jetson Orin Nano 2, a module delivering 78 trillion operations per second of AI compute with 8GB of memory and an eight-core Arm CPU, in the same compact form factor as its predecessor. The company credits improved Tensor Cores and higher memory bandwidth for roughly doubling inference performance relative to the Jetson Orin Nano Super. In its 15-watt mode, the new module draws 40% less power for equivalent peak performance.

The significance is less about the headline number than about what it makes locally feasible. Vision-language models and small language models are the interface layer that lets a robot take an instruction in plain language and ground it in what its cameras see. Running those models over a network link is unacceptable for a drone in flight or a robot in a warehouse aisle, where latency and connectivity failures translate directly into unsafe behavior. Moving that inference on-board, inside a power budget a battery can sustain, is the enabling condition for a large class of products.

“This puts that breakthrough within reach of millions of developers, delivering the performance and energy efficiency needed for real-time reasoning in smart drones, robots, and vision AI systems,” said Deepu Talla, NVIDIA’s vice president of robotics and edge AI. NVIDIA says more than three million developers now build on its robotics stack, and names Wing, Matic Robots, and Cognex among the companies evaluating or using the platform.

The catch is timing: the Orin Nano 2 will not be available until the first half of 2027. For developers, that makes this an architectural announcement rather than a purchasing decision — a signal about what to design for, roughly eighteen months out. Given how much of the current generation of robotics products is constrained by exactly this ceiling, the pre-announcement is arguably as useful as the part.

Source: The Robot Report

Bedrock Robotics Puts Excavators to Work With Nobody in the Cab

Bedrock Robotics announced its first operator-free excavator deployments on active commercial construction sites. The machines are working at a water treatment facility in Nevada with Sundt Construction, on a multi-million cubic yard earthwork project in Texas with Champion Site Prep, and on a 1.2 million cubic yard civil sitework project with Zachry Construction Corp. In each case the excavator is a conventional machine that Bedrock has retrofitted, not a purpose-built robot.

The retrofit is the product. The Bedrock Operator system is a sensor and compute suite that installs in a same-day process requiring no permanent modification to the equipment — meaning a contractor’s existing fleet can be converted and, in principle, converted back. The autonomy model was trained on tens of thousands of hours of field data and handles perception, motion planning, and task execution for clearing, cut-and-fill, and foundation preparation: the repetitive bulk-earthmoving work that consumes the most machine hours on a large site.

Bedrock is explicit that this is not unsupervised autonomy. The machines self-monitor for progress and alert a remote operator when they get stuck, with a person in the loop overseeing operations from off-site. The company frames the goal as minimizing onsite interruptions rather than eliminating human involvement — a framing that has held up better in practice than full-autonomy claims across most of the mobile robotics industry.

The company, based in San Francisco and founded by CEO Boris Sofman, has raised $350 million and is led largely by veterans of Waymo’s autonomy program. That lineage shows in the deployment posture: a full year of testing under human supervision preceded these first operator-free jobs. Construction is an appealing target for that methodology because sites are private property with controlled access, which sidesteps much of the public-road regulatory problem that slowed autonomous vehicles.

Source: The Robot Report

Gatik Raises $200 Million on 85,000 Completed Driverless Deliveries

Gatik closed a $200 million Series D co-led by the Qatar Investment Authority and Koch Disruptive Technologies, with Millennium Management, ARK Invest, and Intact Private Capital participating. The company operates autonomous box trucks on the “middle mile” — high-frequency, fixed regional routes between distribution centers and retail stores, rather than long-haul freight or last-mile delivery.

The operating numbers are what distinguish this round. Gatik reports 85,000 fully driverless orders completed, dozens of driverless trucks running today across North America, and 99% on-time performance, serving PepsiCo, Kroger, Tyson Foods, and Georgia-Pacific across Texas, Arizona, and Arkansas. The company also cites $600 million in contracted revenue — a backlog figure, not recognized revenue, but one that reflects customers committing to volume rather than piloting.

Gatik’s strategic bet has been that route structure matters more than technical generality. A fixed route between the same two facilities, driven dozens of times a day, can be mapped exhaustively, validated statistically, and improved with every repetition. That is a fundamentally easier problem than driving anywhere, and it happens to match a real and large logistics need: retail replenishment runs that are too short for long-haul economics and too frequent to staff comfortably.

The company plans to scale from dozens of trucks to thousands. That transition is where middle-mile autonomy will actually be tested — not on whether the driving works, but on whether fleet operations, maintenance, remote support, and route onboarding scale at anything like the same rate as the software.

Source: The Robot Report

Carbon Robotics Builds a Plant Foundation Model Farmers Can Retune in Minutes

Carbon Robotics, which makes laser weeding equipment, announced a partnership with the data-annotation firm iMerit to build a “large plant model” — a foundation model pre-trained on 150 million labeled plant images gathered globally. The point of the model is not raw classification accuracy but transferability: a system that has seen enough plant structure across regions and species can be pointed at an unfamiliar crop without being retrained.

The user-facing mechanism is unusually simple. A farmer opens an iPad app, reviews thumbnails from their own field, and tags a small number of them as “crop” or “weed.” The model adjusts its behavior immediately — no software download, no retraining cycle, no visit from an engineer. As CTO Alex Sergeev described it, the foundation model “gives you a way to do comparison between this plant and all the plants that the farmer told you ‘These are crops’ and ‘These are weeds.’” Moving from carrots in Arizona to lettuce becomes minutes of configuration rather than a development project.

The scale of the annotation effort is the part worth dwelling on. Carbon had self-labeled roughly 2,000 images before the partnership; iMerit processed approximately one million. That is a five-hundred-fold increase in curated training data, and it came with custom labeling tooling developed collaboratively to make the workflow tractable. This is the unglamorous work that determines whether a field-deployed model generalizes, and it is almost never the part that gets announced.

Carbon also released Carbon ATK, an autonomy kit that retrofits John Deere tractors from model year 2019 onward for fully autonomous weeding with no driver. Like Bedrock’s excavator system, it converts capital equipment farmers already own rather than asking them to buy a new machine — a distribution strategy that is becoming the default in agricultural and construction robotics for good reason.

Source: The Robot Report

EXL Acquires iMerit, Betting That Physical AI Is Bottlenecked on Data Quality

Two days after iMerit’s role in Carbon Robotics’ plant model became public, EXL announced it was acquiring the company. Terms were not disclosed. iMerit provides data annotation and model training services for robotics, autonomous vehicles, and healthcare AI, and operates Ango Hub, a platform for collaborating on complex multimodal data to produce curated and validated training artifacts.

EXL’s rationale is that enterprise AI outcomes now depend more on data quality than on model capability. “Organizations need to bridge the gap between innovation and real-world production, turning AI potential into measurable business results that are trustworthy,” said CEO Rohit Kapoor, framing the deal as combining iMerit’s expert-led training work with EXL’s enterprise data and domain experience.

For robotics specifically, the argument is sharper than it is for text-based AI. Radha Basu, who leads iMerit, put it directly: physical AI must interpret noisy multimodal inputs, reason in real time, and act safely in unpredictable environments. Getting that right depends on human experts identifying edge cases, validating safety-critical behavior, and labeling the situations a model will encounter rarely and must nonetheless handle correctly. Those are judgments that cannot be crowdsourced cheaply or synthesized reliably.

The transaction is a useful market signal. When a services firm specializing in annotation becomes an acquisition target for an enterprise data company, it suggests the buyer expects curated physical-world training data to be a durable, defensible asset rather than a commodity input that model improvements will eventually make unnecessary. The industry has been predicting for years that synthetic data would make annotation obsolete. This deal is a bet against that.

Source: The Robot Report

NSF Funds a Five-Year Center on How People and Robots Adapt to Each Other

The National Science Foundation announced $90 million across three new Science and Technology Centers, each receiving $6 million annually for five years with the opportunity to compete for a further five. One of the three is the Center for Human and Robot Co-Adaptation, led by the University of Texas at Austin. The other two are TEMPEST at Michigan State University, focused on multi-physics and turbulence, and GENIE at Northwestern University, focused on genome intelligence engineering.

The robotics center’s framing is notable for what it does not emphasize. Its stated subject is how people and robots can securely and effectively adapt to one another as service and assistive robots become common in homes, hospitals, workplaces, and public spaces — integrating robotics, AI, and human factors so that robots learn from people, understand their needs, and adjust to different environments and individuals. The research target is the relationship, not the machine.

That emphasis reflects where assistive robotics actually fails. Manipulation and navigation in home environments remain difficult, but the more persistent obstacles are that a robot’s model of what a particular person wants is thin, that the person’s model of what the robot can do is wrong, and that neither updates well over weeks of shared use. Co-adaptation names the problem of both parties learning simultaneously — which is harder than either learning alone, because each is a moving target for the other.

Five-year center-scale funding with a five-year renewal option matters here in a way that grant-scale funding would not. Longitudinal studies of people living with assistive robots take years to run and produce results that no single lab can generate. This is the kind of work that only happens when the funding horizon is longer than a graduate student’s tenure, and it is squarely aimed at supporting independent living and healthcare services rather than industrial throughput.

Source: The Robot Report

The FAA Doubles Its BEYOND Drone Program and Aims It at Heavier Cargo

Transportation Secretary Sean P. Duffy announced that the FAA is launching Phase 2 of its BEYOND program, expanding efforts to integrate drones into the national airspace system. The agency will roughly double the program’s size by selecting up to eight additional state, local, tribal, and territorial entities as lead participants, substantially broadening where advanced drone operations are conducted and studied.

The shift in technical focus is the substantive part. Phase 1 concentrated on smaller aircraft carrying light packages — the familiar suburban delivery-drone use case. Phase 2 targets larger drones hauling cargo to remote and rural areas, and flights above the 400-foot ceiling that has bounded most routine operations, including flights that can gather weather data en route. Both changes push into airspace and payload classes where the safety analysis is genuinely different and where existing operational data is thin.

The rural cargo emphasis is where the near-term public benefit is most legible. Communities without reliable road access pay a large premium in time and cost for medical supplies, parts, and perishables. Aircraft sized for meaningful payloads over long distances change that calculation in a way that shoebox-scale delivery in dense suburbs does not, and the economics are more favorable precisely where ground infrastructure is worst.

Expanding a test-site program is, by design, a data-gathering exercise rather than a regulatory decision. It generates the operational record that eventually justifies routine approvals, and the value of doubling the participant pool lies in the diversity of terrain, weather, and airspace conditions the added sites bring. Whether that record converts into standardized rules on a useful timeline remains the sector’s central open question.

Source: U.S. Department of Transportation

Locus Robotics Acquires a Gripper Startup to Attack the Last 30% of Picking

Locus Robotics, whose autonomous mobile robots have facilitated more than six billion warehouse picks, acquired Nexera Robotics and is integrating its NeuraGrasp soft gripper technology into Array, the company’s mobile manipulator platform. Array combines vision-guided picking with hybrid suction-and-pinch grasping designed for items that defeat pure suction — polybags, cloth, and irregular soft goods.

The acquisition rests on a specific and well-quantified gap. Roy Belak, Locus’s SVP of robotic grasping, estimates that suction handles roughly 60 to 70 percent of warehouse picking tasks; the remaining 30 to 40 percent requires something else. That residual is not a rounding error in a facility processing hundreds of thousands of units a day, and it is the reason each-picking has stayed stubbornly human across most of the industry despite a decade of effort.

Belak also identified tactile sensing as the critical and hardest capability — and gave a pointed reason why. Simulation, which has driven most of the recent progress in robot learning, does not adequately capture contact physics. You cannot generate reliable touch data in a simulator the way you can generate visual data; deformable objects, friction, and slip behave in ways that simulators approximate poorly. That forces the data collection back into the physical world, which is slow and expensive — the same bottleneck Carbon Robotics and iMerit ran into from a different direction.

Locus is deliberately optimizing for reliability and pick quality rather than speed, emphasizing damage avoidance, double-pick prevention, and accuracy over raw throughput. This is a mature position. A picking robot that is fast but occasionally grabs two items or crushes one imposes downstream costs — mis-shipments, returns, manual audits — that erase the labor savings. Array is already in early customer deployment with its original gripper, with the Nexera technology slated for future integration.

Source: The Robot Report

A Soft Robot That Leaps Indefinitely on Light Alone, With No Reset Between Jumps

Researchers at NC State University published a soft robot that jumps continuously under infrared illumination, requiring no motor, no battery, no onboard control, and — most unusually — no manual reset between jumps. The work appeared in PNAS on August 27 under the title “A Self-Resetting Soft Ring for Autonomous, Continuous Leaping in Unstructured Environments,” authored by Fangjie Qi, Caizhi Zhou, Haitao Qing, Haoze Sun, Yaoye Hong, and Jie Yin.

The construction is almost startlingly simple: a teardrop-shaped ribbon of liquid crystal elastomer with a thin V-shaped aluminum tube at one end. Under infrared light, the ribbon’s illuminated surface contracts, which makes it rotate. The rigid aluminum component prevents the structure from simply rolling, so instead it twists tighter and tighter, storing elastic energy until the accumulated strain releases against the ground and throws the robot into the air. The geometry then returns it to a configuration where the cycle begins again — which is what “self-resetting” means here, and it is the contribution.

The performance is substantial for a device with no actuator. The robots jump more than 80 body heights vertically and leap over three body lengths forward, and they traversed grass, sand, rocks, mulch, water surfaces, slopes, and hurdles in proof-of-concept testing. Control, such as it is, comes from light intensity: too weak and the robot does not jump, too strong and it jumps erratically in unpredictable directions.

The researchers are candid that there are no immediate applications. The value is in the mechanism. Nearly every mobile robot in the other nine items this week carries its own power, computation, and actuation, and most of its mass and cost goes to those three things. A machine that offloads all of them to an external light field is a different design point entirely — one plausibly relevant to swarms of very small robots, to environmental sensing in places where batteries are impractical, and to locomotion across terrain that wheels and legs handle badly. Fundamental mechanisms like this take a long time to reach products, and they are how the field acquires genuinely new options rather than better versions of existing ones.

Source: NC State University News