Robotics Weekly Review 2026-08-22
Week In Review
If there was a single throughline in robotics this week, it was the sound of an industry building its supply chain. The most-watched event was financial: Unitree Robotics went public on Shanghai’s STAR Market, and its first-day surge made it, by its own account, the first major legged-robotics company to list (What does Unitree Robotics’ IPO mean for the humanoid industry?). But the more consequential announcements sat one layer beneath the robots themselves. Schaeffler said it will mass-produce formed strain wave gearboxes for humanoids starting in 2027 (Schaeffler plans to mass produce gearboxes for humanoid robots in 2027), attacking a component that it says accounts for roughly half of a humanoid’s build cost. In South Korea, a fifty-year-old materials company bought a Brooklyn e-textiles startup to get into robot skin (Unichem acquires Loomia to accelerate entry into the humanoid ‘skin’ market). Actuators and touch sensors are unglamorous, and they are exactly what determines whether humanoids stay a demo genre or become a manufactured product.
The learning layer advanced in parallel, and in a direction that matters for every item below. Generalist AI released GEN-1.5, reporting that one-shot and few-shot learning of physical skills emerged from scale rather than from a training objective aimed at producing them (GEN-1.5: Embodied Foundation Models are One-Shot Learners). A robot that can pick up a new task from a twelve-second demonstration is a different economic proposition from one that needs a data-collection campaign per skill. The same shift shows up in applied form elsewhere: Gravis Robotics raised $200 million to put autonomous control on existing excavators, having trained largely in simulation (Gravis Robotics raises $200M for autonomous construction), and Diligent Robotics built Moxi 2.0 around a hospital-specific world model refined by five years of field data (Five years of operation shape Diligent Robotics rollout of Moxi 2.0).
Deployment news clustered around unromantic, high-volume work. Pudu launched an autonomous pallet handler aimed at crowded loading docks and production floors (Pudu Robotics launches new MP2000 autonomous forklift). Amazon said it intends to reach nearly 500 U.S. cities and towns with drone delivery by the end of this year (Amazon plans to expand Prime Air to nearly 500 cities by the end of 2026). Serve Robotics added Grubhub to its sidewalk-delivery marketplace partners (Serve Robotics to deploy its autonomous delivery robots with Grubhub) — notable partly because Serve now also owns Diligent, meaning one company is running both restaurant sidewalks and hospital corridors on shared fleet-learning infrastructure. None of these are humanoids. All of them are robots doing repetitive transport that someone is paying for today.
The week also supplied its own counterweight. Analysts quoted in the Unitree coverage noted that fewer than 10% of the company’s sales appear to be commercial deployments, with most units going to research labs and lobbies — a reminder that unit shipments and useful work are different metrics. The most human-centered item of the week made the point from the other direction: ATDev’s ARPA-H-funded work on an autonomous wheelchair with an integrated arm (ATDev gives update on its journey building autonomous wheelchairs) is deliberately not chasing full autonomy, and is designed around a “spectrum of autonomy” that keeps the user in control. Between a foundation model that learns in seconds and a wheelchair built by the people who will sit in it, the field’s two halves — capability and trust — were both visibly moving.
Items
Generalist AI’s GEN-1.5 Learns New Physical Tasks From a Single Demonstration
Generalist AI published GEN-1.5 on August 19, and the headline claim is unusual enough to be worth stating precisely: the model can acquire a new manipulation task from a single demonstration, in context, without any gradient updates or fine-tuning. In machine-learning terms, this is the same trick large language models perform when you paste an example into a prompt and they follow the pattern. Doing it with physical actions — where the “prompt” is a few seconds of video showing hands moving objects — has been a long-standing goal precisely because it removes the field’s dominant cost, which is collecting demonstration data for every task you want a robot to perform.
The reported numbers give a sense of both the promise and the current ceiling. Across ten short tasks, GEN-1.5 averaged 59% success from a single 3-to-12-second demonstration with no weight changes at all. Allowing ten gradient steps on about five minutes of data — roughly fifty demonstrations — raised the average to 83%. Even one gradient step on a single minute of data reached 66.5% on a held-out task. Generalist notes that fine-tuning shifts the model’s weights by less than 0.15%, which supports their interpretation that adaptation is reconfiguring capabilities the model already has rather than teaching it something new.
Architecturally, GEN-1.5 is a large multimodal model that takes video with a 30-second memory window alongside language, proprioceptive, and other sensor inputs, and emits action trajectories at 100 Hz. That output rate matters: it is fast enough for closed-loop control of real hardware rather than for issuing high-level plans that some other controller must execute.
Two generalization results stand out beyond the benchmark averages. The model can sometimes learn across the embodiment gap — a human demonstrates a task with their own hands, visible through the robot’s cameras, and the robot reproduces it immediately afterward, despite having entirely different actuators. It can also transfer demonstrations performed in simulation to real hardware zero-shot. Generalist attributes the emergence of these behaviors primarily to scale, describing GEN-1.5 as the product of more than eight months of continuous pretraining on physical-interaction data plus targeted architectural and algorithmic changes, rather than a redesign aimed at one-shot learning.
The company is careful about scope, and readers should be too: the tasks demonstrated are short-horizon and individually simple. What is significant is not that a robot stacked a block, but that the capability to learn which block, from one example, appeared without being explicitly trained for.
Source: Generalist AI
Unitree Robotics Goes Public on Shanghai’s STAR Market
Unitree Robotics listed on Shanghai’s STAR Market this month and closed its first day of trading 460% above its opening price, raising 6.1 billion yuan — roughly $905 million. The Hangzhou company describes itself as the first major legged-robotics firm to go public, and on the numbers it has a reasonable claim to being the first to demonstrate that selling legged robots can be a profitable business rather than a research subsidy.
Its financial trajectory is genuinely steep. Unitree reported approximately $41 million in profit for 2025, up from $14 million in 2024, following a $1.6 million loss in 2023. Unit sales followed a similar curve: from five humanoids sold in 2023 to 5,215 in 2025, a figure that plausibly makes it the global volume leader in humanoid robots. For a category that spent a decade as a demonstration medium, a profitable manufacturer shipping thousands of units annually is a real structural change.
The industry reaction captured in The Robot Report’s coverage was notably measured, and the caveats are worth taking seriously. Gavin Kenneally of Ghost Robotics estimated that fewer than 10% of Unitree’s sales go to commercial deployments, with the large majority landing in research facilities, universities, and corporate lobbies. Nicolaus Radford of Persona AI argued the company has not yet demonstrated operational durability or repeat business at industrial scale. Both critiques point at the same gap: shipping a robot and having that robot do economically useful work every day are separate achievements, and public-market valuations tend to price the first as though it implied the second.
A geopolitical constraint also sits on the company’s growth path. The FCC has restricted U.S. imports of certain foreign-made robots, which could limit Unitree’s access to the American market regardless of its cost advantages. For a manufacturer whose competitive position rests substantially on price, losing a major market is more than a rounding error.
Still, the listing changes the sector’s information environment in a useful way. A public humanoid manufacturer must file audited financials, which means the field will finally have a reliable, recurring dataset on what it costs to build these machines and what customers actually pay — replacing a decade of vendor-supplied claims with something an analyst can check.
Source: The Robot Report
Schaeffler Will Mass-Produce Humanoid Gearboxes Using Forming Instead of Machining
Schaeffler Technologies announced plans to begin mass production of formed strain wave gearboxes for humanoid robots in 2027, targeting what may be the single largest cost line in a humanoid’s bill of materials. By Schaeffler’s accounting, actuator gearboxes represent roughly half of a humanoid robot’s manufacturing cost — which means that a company solving gearbox economics has more leverage over humanoid adoption than most of the companies building humanoids.
Strain wave gearboxes, sometimes called harmonic drives, are the standard solution for robot joints because they deliver very high gear reduction with minimal backlash in a compact, lightweight package. Those properties are what let a robot arm hold a position precisely under load. The problem has always been manufacturing: the flexible spline at the heart of the device is conventionally produced by precision machining, a slow and capital-intensive process that resists scaling.
Schaeffler’s contribution is a process change rather than a design change. Instead of cutting the component, the company forms it, using high pressing forces to shape the part in seconds rather than minutes. The reported results are a 25% reduction in cost and a 75% reduction in material consumption, at performance the company says is comparable to machined equivalents. The material savings figure is the striking one: forming displaces metal rather than removing it, so most of what would have become swarf stays in the part.
Schaeffler is not new to this. The company says it has supplied more than two million formed strain wave gearboxes to markets worldwide over the past decade, largely through automotive applications — which is to say the process is production-proven at volume, and the 2027 announcement is about adapting a mature manufacturing capability to a new specification rather than commercializing a laboratory result. Production will begin in Germany, with expansion to other regions planned afterward.
The broader significance is about bottlenecks. Every projection of humanoid robots reaching meaningful deployment numbers implicitly assumes someone can produce tens of millions of precision actuators annually. Very few organizations have ever done precision mechanical manufacturing at that scale; most of them are automotive suppliers. This week’s announcement is a sign that tier-one automotive suppliers have decided the humanoid market is real enough to retool for.
Source: The Robot Report
Unichem Acquires Loomia, Betting That Robots Need Skin
South Korea’s Uni-Chem, a materials company founded in 1976 that has spent most of its life making leather for automotive interiors and fashion, acquired the Brooklyn flexible-electronics startup Loomia Technologies for approximately $6 million (8.6 billion won). The structure — $4 million cash, $1 million in Unichem stock, and $1 million in performance earn-outs, closing October 15 via reverse triangular merger — is modest by robotics-deal standards. The strategic logic is not.
Loomia, founded in 2014 by Stanford graduate Madison Maxey, developed what it calls the Loomia Electronic Layer, or LEL: a technique for bonding electronic circuits directly to flexible materials such as cloth and leather, escaping the geometric constraints of rigid printed circuit boards. The sensing layer combines capacitive sensors and force-sensitive resistors, and — importantly for manipulation — detects shear force as well as vertical pressure.
That distinction is more than a specification detail. Knowing how hard you are pressing on an object tells you whether you have contact. Knowing whether the object is sliding sideways against your fingers tells you whether you are about to drop it. Human grip relies heavily on shear-sensing to modulate force in real time, which is why people can hold an egg and a hammer with the same hand without thinking about it. Vision cannot supply this information; the contact patch is precisely the part of the scene the hand is occluding. As Loomia’s team framed it, tactile sensing takes robotics “beyond just vision.”
The commercial arrangements suggest both parties think the timing is right. Manufacturing will run through R&Y, an automotive supplier with over $1 billion in revenue, which plans to deliver engineering samples by January 2027 and is targeting Tesla, OpenAI, Boston Dynamics, 1X Technologies, and Meta as customers. Volkswagen has already selected Loomia’s technology for next-generation seat heating — a reminder that the same flexible-electronics substrate has a non-speculative automotive business underneath the robotics ambitions. Maxey and Unichem chief science officer Kim Han-joo will serve as co-CEOs. Loomia’s tactile developer kit, launched last year, is currently sold out.
Source: The Robot Report
Amazon Aims to Bring Prime Air Drone Delivery to Nearly 500 Cities This Year
Amazon announced plans to expand Prime Air to nearly 500 U.S. cities and towns by the end of 2026 — roughly a sixfold increase over its current service footprint. The service operates today across seven states: Arizona, Florida, Kansas, Louisiana, Michigan, Nebraska, and Texas. Imminent launches were named for Chicago, Cleveland, Atlanta, Syracuse, and Boise, with more communities to follow later in the year. Amazon also recently opened its first international Prime Air location in Darlington, U.K.
The operational envelope explains why this is expandable rather than exotic. Prime Air carries packages of five pounds or less that fit in a large shoebox, within a 7.5-mile radius of a fulfillment center. Amazon says more than 60% of its frequently purchased items qualify. Delivery lands in as little as 30 minutes, with most arriving around an hour after checkout. These are deliberately unambitious parameters, and that is the point: the company is not trying to replace trucks, it is skimming the specific subset of orders where a small aircraft flying a straight line beats a van navigating streets.
The regulatory position is what makes the scale-up plausible. Prime Air holds FAA Part 135 certification — the same framework applied to commercial air carriers, and the highest level of FAA oversight available for drone delivery. Earning it is slow and expensive; holding it means expansion becomes a question of adding approved operating areas rather than relitigating the safety case each time. The aircraft carry onboard cameras and sensors for navigation and obstacle detection, plus a Detect-and-Avoid system that monitors surrounding airspace and makes routing decisions in flight.
Amazon reported delivering hundreds of thousands of packages by drone during 2026 so far. That is a meaningful volume for an industry that spent years in pilot programs, and a small one against Amazon’s total shipment count — which is a fair description of where drone delivery sits generally. “Customers already turn to Amazon for fast same- and next-day delivery, and Prime Air provides them an even speedier option,” said David Carbon, vice president of Amazon Prime Air.
The interesting question the announcement raises is about ground infrastructure rather than aircraft. A 7.5-mile service radius tied to fulfillment centers means drone coverage is a function of warehouse density, so reaching 500 communities is partly an aviation problem and substantially a real-estate one.
Source: The Robot Report
Serve Robotics Adds Grubhub, Expanding Sidewalk Delivery Across Three Cities
Serve Robotics announced a partnership with Grubhub — now a Wonder subsidiary — to offer autonomous sidewalk delivery through the Grubhub marketplace. The launch covers over 100 participating merchants in Chicago, nearly 200 restaurants in Los Angeles, and Wonder’s location in Alexandria, Virginia. Serve simultaneously launched with DoorDash in Washington, D.C. and San Jose, California.
The multi-marketplace pattern is the strategically interesting part. Serve is not building a consumer app and competing for demand; it is positioning itself as delivery capacity that any marketplace can call. Adding Grubhub while expanding with DoorDash means the same physical fleet in a given city can be fed orders from multiple demand sources — which matters enormously for a business whose unit economics depend on robot utilization. A sidewalk robot idle between orders is pure cost.
Serve’s Q2 2026 financials show the model gaining traction, though from a small base: $3.2 million in revenue, up 9% sequentially and 404% year over year. DoorDash-sourced revenue grew nearly 50% sequentially. More than half of total revenue is now recurring. Perhaps the most unexpected line is advertising, which accounts for nearly half of food-delivery revenue — the robots carry display surfaces, and it turns out a slow-moving vehicle at eye level on a busy sidewalk is a saleable advertising medium. That may prove to be a durable structural advantage over delivery methods that lack a billboard.
The company also launched microdepots in Miami and introduced two new products: Beacon, a countertop device for restaurants, and Characters, branded conversational AI robots. “Every new partner puts more robots to work, and every delivery makes the whole fleet smarter,” said CEO Ali Kashani — a claim about fleet learning that connects directly to Serve’s other robotics line. Having acquired Diligent Robotics in January, Serve now operates both restaurant sidewalk delivery and hospital-corridor logistics, two environments that differ enormously in stakes but share the underlying problem of navigating crowded human spaces reliably.
Source: The Robot Report
Pudu’s MP2000 Takes Autonomous Pallet Handling Into Crowded Facilities
Pudu Robotics launched the MP2000, an AI-driven autonomous pallet handler built for high-frequency, heavy-payload transport in crowded manufacturing and logistics environments. The company reports pallet pickup cycles as fast as 20 seconds — the metric that matters most in this category, since an autonomous forklift that is safe but slow simply loses to a human driver on throughput.
Sensing is handled by 3D lidar and depth cameras, which support real-time path adjustment and obstacle avoidance between pickup and drop-off. The perception system also validates pallet dimensions and compensates for misalignment, which addresses the mundane reality that defeats many warehouse automation projects: pallets in the real world are rarely squared up, frequently damaged, and often not quite the shape the spec sheet assumed. Pudu says the MP2000 handles three-runner pallets, perimeter-base designs, and customized non-standard load carriers — an acknowledgment, in the company’s framing, that pallet handling rarely follows a single standard.
The fleet software is designed around graceful degradation. Distributed fleet intelligence coordinates robots both locally and fleet-wide, and the system is built to keep operating through network outages rather than halting — a meaningful design choice in facilities where industrial Wi-Fi coverage is imperfect and a stalled fleet blocks aisles. The robots also interface with building infrastructure such as automatic doors and elevators, allowing floor-to-floor transport without human intervention.
One pragmatic feature deserves mention: a manual override converts the MP2000 into a conventional powered forklift. This is the kind of detail that distinguishes equipment designed for real facilities from equipment designed for demonstrations. When autonomy fails at 2 a.m., the alternative is not a support ticket — it is an operator who needs the pallet moved now.
The MP2000 extends Pudu’s existing T-Series autonomous mobile robots upward into heavier palletized loads. It targets loading docks, warehouses, line-side storage, and production workshops. Pricing and delivery timelines were not announced.
Source: The Robot Report
Gravis Robotics Raises $200 Million to Automate Excavators
Gravis Robotics announced a $200 million Series A from SoftBank on August 17, which the company describes as the largest funding round in construction robotics history. The Zurich-based company, founded in 2022 as a spinout from ETH Zurich, develops physical AI software for autonomous earthmoving equipment. Valuation was not disclosed.
The company’s approach avoids the trap that has caught several construction-robotics predecessors: it does not build machines. Gravis retrofits existing excavators and construction equipment with the Gravis Rack, an autonomous control kit and software stack that is deliberately fleet-agnostic — compatible with machinery from Caterpillar, Case, Develon, John Deere, JCB, Hitachi, Sumitomo, Yanmar, and Volvo, among others. Contractors already own billions of dollars of iron that works fine; selling them a control upgrade is a far easier proposition than selling them a new fleet.
Earthmoving is a harder autonomy problem than its rough appearance suggests. Soil is not a rigid body. Resistance on a bucket depends on moisture, compaction, embedded rock, and how the previous pass disturbed the material — so the machine must sense and respond to forces it cannot see in advance. Gravis addresses this through large-scale synthetic training, describing simulation across billions of cubic yards of virtual material. Co-founder Dominic Jud characterizes the system as grounding physical input “in machine telemetry, responding to varying subterranean forces at microsecond speeds.” The company claims productivity improvements of up to 30% over peak manual operation — notable partly because peak manual operation is a demanding baseline, given that skilled excavator operators are very good.
Gravis has deployed across four continents and was selected to lead an $8 million U.K. government-backed CAM Pathfinder project with Flannery Plant Hire, the country’s largest equipment rental provider. The rental channel is a shrewd distribution path: rental firms buy equipment in volume, run it across many job sites, and have direct financial exposure to utilization rates, making them unusually motivated early adopters for productivity technology.
Source: The Robot Report
Diligent Robotics Rolls Out Moxi 2.0, Shaped by Five Years in Hospitals
Diligent Robotics began deploying Moxi 2.0 in August, the successor to a hospital logistics robot that has now accumulated five years of operational experience across more than 25 hospitals. The platform was unveiled in October 2025; the company was acquired by Serve Robotics in January 2026 for $29 million. What makes this rollout worth attention is less the hardware than the fact that its design brief was written by half a decade of field data rather than by a product team’s assumptions.
The performance upgrades are substantial. Moxi 2.0 carries 10 times the onboard compute of its predecessor, with perception running 10 to 15 times faster — the difference between a robot that pauses to think in a busy corridor and one that keeps moving. Runtime extends to 18 hours a day, split into nine-hour sessions with 30% faster charging, which aligns the robot’s availability with hospital shift patterns instead of forcing workflows around charging gaps. Improved edge-case recovery lets the robot resolve more unexpected situations on its own rather than stopping and requesting help.
Underneath sits what Diligent calls a Robotic World Model, purpose-built for hospital complexity, strengthening navigation and task completion as deployment data feeds back into training. The pipeline runs on NVIDIA Isaac Sim for simulation and AWS SageMaker HyperPod for cloud model training. This is the same learning-flywheel argument Serve makes about its delivery fleet, applied to an environment with considerably less tolerance for error.
The most telling changes are the least technical. Upgraded cameras and an expanded sensor suite arrived alongside redesigned storage drawers and reshaped handles — the latter informed directly by nurse and pharmacy staff feedback. Five years of watching people use a robot apparently produced as many insights about drawer ergonomics as about autonomy stacks, which is an honest reflection of what deployed robotics actually involves.
Early Moxi 2.0 sites include Endeavor Health Edward Hospital, Providence Saint John’s Health Center, and Children’s Hospital Los Angeles, with nationwide expansion continuing. The Los Angeles facility has documented over 40,000 completed deliveries, which Diligent frames as roughly 16,000 hours of fetch-and-carry work that clinical staff did not have to do. That hospital has grown its fleet from two robots to three, with utilization rising over 10% in the second quarter alone — the kind of organic expansion that suggests the value is being felt by the people doing the work.
Source: The Robot Report
ATDev Is Rebuilding the Power Wheelchair as a Robot
ATDev, based in Orange, California, provided an update on the Robotic Assistive Mobility and Manipulation Platform, or RAMMP — a power wheelchair with an integrated robotic arm, funded by the Advanced Research Projects Agency for Health and led by the University of Pittsburgh. The goal is to let users perform reaching, lifting, eating, drinking, and door-opening tasks they cannot currently do independently. There are roughly 3.5 million power wheelchair users in the United States.
ATDev’s decision to redesign the entire platform from scratch rather than bolt an arm onto an existing chair is the technically significant choice here. Power wheelchairs are medical devices built on power electronics and software architectures that predate modern robotics by decades; adding an autonomous arm to that foundation means inheriting every constraint of it. Rebuilding is slower and far more expensive, and it is the only path to a chair where mobility and manipulation are coordinated rather than merely co-located.
The control philosophy is what distinguishes this from an autonomous vehicle in wheelchair form. ATDev describes a “spectrum of autonomy” in which the user chooses, in real time, between fully autonomous operation and direct control of individual motors. The system uses shared autonomy: the chair proposes actions and the user supplies corrective input when it gets something wrong, rather than the chair operating independently and the user hoping it does not. For a device carrying a person who may have complete paralysis, this is a safety architecture as much as a user-interface decision — full autonomy would require a level of guaranteed correctness that no current robot can offer.
The development process reflects the same principle. Wheelchair users serve as researchers and principal investigators on the program, not as test subjects consulted after design decisions are made. The team is also exploring teleoperation, so that a user confined to bed could send the chair to fetch something from another room — a capability that inverts the usual assumption that the chair’s value is limited to moments when the user is sitting in it.
Longer term, the program includes federated learning, letting individual chairs share learned behaviors through what ATDev describes as an app-store model: a task one user teaches their chair could become available to others. The company expects FDA approval and insurance reimbursement to take about five years, with no launch date announced. Reimbursement, not robotics, will likely determine whether this reaches the people who need it.
Source: The Robot Report