News
What's happening in robotics and AI — curated by the wujec.ai editors.
The first excavators working with nobody in the cab are not new machines — they are other people's diggers with a kit bolted on
Bedrock Robotics says its excavators are now working fully autonomously on active commercial job sites — sharing those sites with human crews and human-driven equipment rather than running on a closed test yard. Trade coverage put the count at three major sites. The detail that matters is what the company does not sell. Bedrock builds no excavator. The Bedrock Operator is a retrofit kit: sensors, an in-cab computer and drive-by-wire control that go onto a machine the contractor already owns. The firm states it fits 20 to 80 tonne excavators from the usual makers, mounts in a single work shift and leaves no permanent modification. That removes the obstacle that has slowed every previous attempt at construction autonomy — nobody has to replace a fleet, and nobody has to wait for Caterpillar, Deere or Komatsu to build an autonomous model. Bedrock explains why it began with the hardest machine rather than the easiest. The excavator is roughly a quarter of a typical fleet and has more degrees of freedom than almost anything else on site; a system built for that case extends down to dozers and loaders, while the reverse does not. The stated destination is a fleet that sequences its own work. The autonomy has a clear edge, and the company draws it honestly: the robot digs, swings and loads dump trucks that people still drive. Remote human supervisors can intervene. Bedrock says the safety system was tested in the field for two years before the operator left the cab. The business case is a labour shortage, not a novelty. More than 40 percent of the American construction workforce is expected to retire within five years, and the contractors Bedrock quotes describe turning down work for lack of operators. One project manager frames it differently again: his best operators lose interest during months of repetitive mass excavation, so giving the truck-loading cycle to a machine keeps skilled people on work that needs judgement. About one in five workplace deaths in the United States happens in construction. Bedrock was founded by veterans of Waymo's autonomous trucking programme and raised 270 million dollars in February 2026 at a 1.75 billion dollar valuation. What it still does not publish is a technical data sheet: the hardware specifics circulating for this system come from a market analyst, not from the company.
Bedrock Operator →Fifty humanoids, 63,000 kilometres, 189,000 incidents — a carmaker publishes the first real mileage log of a humanoid fleet
Almost every humanoid announcement ends at the demonstration. GAC Group, the Guangzhou carmaker, has published something rarer: an operating log. In a press release dated 11 August 2026 the company reported that its wheel-legged GoMate Mini has been deployed on seven flagship projects, that close to 50 units are in the field, that the longest-running project has been operating continuously for twelve months, and that the fleet has covered roughly 63,000 kilometres of patrol distance while handling about 189,000 on-site incidents. Divided by the fleet, those numbers describe a working machine rather than a stage prop: roughly 1,260 kilometres and about 3,800 handled incidents per robot. For comparison, most humanoid makers still report their progress in cumulative units built, not in distance travelled or tasks closed — a metric that only becomes available once the machines are somebody's responsibility on a real shift. The deployment route is unusual too. On 31 July 2026 GoMate Mini was formally enrolled into the security system organised by the China Security Association, which puts the robot into service in numbers rather than in pilots. Commercial orders are reported at close to ten million yuan — around 200,000 yuan a unit at the stated fleet size, which is service money, not the price of a robot. The robot behind the log is the fourth generation of the GoMate line. It uses a variable wheel-foot structure with three switchable modes, runs more than eight hours on a charge and can right itself after a fall — the mundane capability that decides whether a night patrol needs a human minder. GAC says it can switch from a standing patrol posture into a carrying configuration, so one machine covers both inspection and material transport. The corporate structure changed with it. GAC incubated the robotics work into a separate company, Huilun Technology, in early 2026, and the same press release announces a funding round of more than 100 million yuan. The pitch is explicitly automotive: the firm says it has carried vehicle-grade production standards, quality control and supply-chain management into robot manufacturing, and that it develops its own axial-flux joint motors, integrated joint modules and rigid-flexible dexterous hands. Capacity planning is stated at the ten-thousand-unit scale. What the log does not say matters as well. GAC does not report how many of the 189,000 incidents required a human to step in, how many robots were withdrawn from service, or what the twelve-month project cost to keep running. Mileage is a measure of exposure, not of autonomy — but it is the first number in this industry that cannot be staged for a conference floor.
GAC GoMate →Amazon says drones will reach nearly 500 US towns this year — that is a sixfold jump in place names, not in launch sites
On 19 August 2026 Amazon said Prime Air would serve "nearly 500 cities and towns" in the United States by the end of the year. Today the service flies from 11 locations in 10 metro areas across seven states — Arizona, Florida, Kansas, Louisiana, Michigan, Nebraska and Texas. New sites were named in the Chicago, Atlanta, Cleveland, Syracuse and Boise metro areas. The two numbers are not measured in the same unit. A Prime Air site works a radius of about 7.5 miles, which is roughly 175 square miles of ground, and an American metropolitan area of that size routinely contains dozens of separately incorporated cities, towns and villages. Adding a handful of sites therefore adds hundreds of place names. "We've already delivered hundreds of thousands of packages to customers by drone this year," said David Carbon, vice president of Prime Air — a figure that, spread over a year and 11 sites, describes a working service rather than a mass one. The aircraft doing the flying is the MK30, in customer service since November 2024. It lifts off on six motors, transitions to wing-borne cruise at up to 73 mph, flies up to 400 feet above ground and returns to base within a 7.5-mile radius. Its maximum take-off weight is 83.2 lb, of which the customer's parcel may be no more than 5 lb — a ceiling Amazon says still covers about 60% of what its customers buy most often. The constraint that decides whether the target is met is not the drone. Prime Air holds an FAA Part 135 air-carrier certificate and flies beyond visual line of sight, but each new site needs its own regulatory clearance and environmental review before a single package leaves the ground. Amazon has published a target measured in towns; what is worth watching between now and December is the number measured in sites.
Amazon Prime Air MK30 →The two biggest drone delivery services agree on the payload and disagree on everything after the flight
Two companies fly more delivery drones than anyone else, on opposite sides of the world, with no shared supplier and no shared regulator. They have converged on the same payload: Amazon's MK30 is certified for 5 lb (2.3 kg), and Meituan's fourth-generation aircraft is quoted at 2.5 kg in China and 2.4 kg in the export version sold under the Keeta Drone brand. Amazon says that ceiling covers roughly 60% of what its customers most often buy. The agreement ends there. Amazon's drone descends over a customer's garden and releases the parcel from a low hover onto the lawn — which works in American suburbs and nowhere else. Meituan's flies a fixed corridor between towers and ends at a pick-up cabinet, because in a high-rise district there is no lawn and no legal way to drop anything. Each design encodes the city it was built for. The consequences show up in the range figures, and they are the reverse of what the technology suggests. Amazon's aircraft covers a 12 km radius, roughly 450 km2 from one site; Meituan's covers 5 km fully loaded. The Chinese drone is the more constrained machine on paper and the busier one in practice: Meituan had passed 900,000 real customer orders by May 2026, flying scheduled routes rather than waiting for demand. That is also why the aircraft Meituan announced on 21 May matters more than its specification sheet suggests. The M-Drone 4L Winch does not land at all: it hovers and lowers the package on a cable, which means a destination no longer needs a cabinet, floor space, power or a landlord's signature. Payload rises to 4.5 kg and range to 20 km fully loaded, and Meituan states the noise was cut specifically so the aircraft can fly at night. Remove the cabinet and the Chinese model stops being a city-centre special case. The remaining difference between the two services is then not the drone but the paperwork: Amazon needs a separate environmental review and authorisation for every new site, while Meituan is recruiting third-party operators to fly its network under a licence issued for the country as a whole.
Meituan M-Drone 4 →Six in ten humanoids shipped this year went to a stage or a data lab — not to a factory
Humanoid robots are shipping in numbers that finally look industrial: more than 22,000 units in the first half of 2026, close to 300% up year on year, according to Counterpoint Research. What those machines are bought for is a different story. Two categories still account for more than 60% of all shipments: entertainment and performance, and data production and research — that is, robots bought to appear in front of an audience, and robots bought to generate training data for the models that will one day operate them. Their combined share slipped moderately against last year, but it did not stop being the majority. The segments usually named as the point of the whole exercise remain small. Intelligent manufacturing rose to 13% of shipments; warehousing and logistics to 5%. The figures match what this catalogue has spent August writing about. The two humanoids that broke the 100 m record in Beijing were built to compete, not to work. EngineAI's T800 is the only machine admitted to a humanoid fight league in which every entrant is given identical hardware. DroidUp's Rena performs kunqu opera. None of that is a criticism of the machines — but it is a long way from the images of humanoids carrying parts down a production line that accompany most coverage of the sector. Counterpoint expects the balance to reverse, with service and industrial applications overtaking performance and data collection over the next five years, and total shipments passing 50,000 units in 2026. For now, the most common job for a new humanoid is to be watched.
Moxi 2.0 starts shipping — and the spec sheet names an NVIDIA chip five years older than the launch page
Diligent Robotics, a Serve Robotics company, announced on 17 August 2026 that Moxi 2.0 — the second generation of the most widely deployed hospital delivery robot in the United States — is now rolling out to health systems, naming Endeavor Health Edward Hospital, Providence Saint John's Health Center and Children's Hospital Los Angeles as the first sites. The release finally puts numbers on hardware that the October 2025 announcement described only in adjectives. Moxi 2.0 runs "up to 9 hours at a time" on a charge, charges 30% faster than its predecessor, and the company claims "up to 18 hours of operating time per day". Those three figures fit together exactly: two full 9-hour shifts leave six hours in the day for two recharges, roughly three hours each — which puts the first generation at about four and a quarter hours to fill the same battery. Perception is stated at "10–15 times faster than the previous generation", the world model was trained on Amazon SageMaker HyperPod, and the perception stack borrows "aspects of the NVIDIA Cosmos open world models such as the 3D lidar tokenizer" — the first confirmation from the maker that the robot carries lidar at all. One line does not match the company's own product page. The shipping release credits "upgraded NVIDIA-powered A2000 compute", while the Moxi 2.0 launch page still says the robot is "powered by NVIDIA Thor" and specifically names IGX Thor, "an NVIDIA Blackwell-powered, industrial-grade" platform. These are not the same silicon: the RTX A2000 is an Ampere-generation professional card launched in August 2021, five years and three architecture generations behind Blackwell. Both claims are live on the company's channels on the day the robot begins shipping, and Diligent has not reconciled them. The "10x the compute of Moxi 1.0" figure is repeated in both, so the multiplier is not in dispute — the part number is. The deployment figures are more solid, because one customer counted them. Children's Hospital Los Angeles reports that its Moxi has completed "more than 40,000 deliveries, representing over 16,000 hours of work", and grew its fleet from two robots to three; utilisation rose more than 10% in the second quarter. Divide those two numbers and the robot finishes a delivery every 24 minutes it is on shift — a useful benchmark for anyone comparing hospital logistics platforms, and a rate no competitor in this niche has published against real hours. This catalogue records the runtime, charging and lidar figures from the shipping release, and notes the compute discrepancy in the Moxi 2.0 profile rather than picking a side.
Moxi 2.0 →Europe's largest robotaxi plan was signed by two companies that will not own a single car — Pony.ai and Uber add 2,000 vehicles, and a Rimac spin-off carries the permits
Pony.ai and Uber announced on 14 August 2026 that they will expand their partnership to more than 2,000 Pony.ai robotaxis across Europe, adding four European cities to the service already running in Zagreb, and covering the Middle East in the same agreement. Target cities and timetables, the companies said, will be announced in phases. The vehicles are Pony.ai's seventh-generation robotaxis, the same kit that entered commercial service in Beijing, Guangzhou and Shenzhen in November 2025 and that already carries paying passengers in Croatia. Pony.ai's founder and chief executive James Peng framed the deal as combining his company's autonomous driving technology and operational experience with Uber's global mobility platform; Sarfraz Maredia, who runs autonomous mobility at Uber, described the goal as moving "from individual launches to repeatable commercial scale". The most instructive part of the announcement is the division of labour, because it is not the usual one. Pony.ai supplies the driving system. Uber supplies the demand, through its app. And Verne — a Croatian company founded by Mate Rimac and spun out of the Rimac Group, with Marko Pejković as chief executive — owns the fleet, runs it on the ground, and leads the regulatory approval process in every launch city. In other words, the two large partners in the biggest robotaxi commitment yet made in Europe will not buy the cars and will not sit across the table from the regulators. That structure is a direct answer to what makes Europe hard. There is no single European authorisation for driverless passenger service; each country grants it separately, on its own terms. A company that wants five cities in Europe needs to run that process up to five times, in five legal systems, and it helps considerably to be a local operator while doing so. Verne says it is in talks with eleven cities in Europe, the United Kingdom and the Middle East, with more than thirty further cities under evaluation. Scale deserves a caveat that the press release does not offer. The Zagreb service, which opened in 2026 as the first commercial robotaxi operation in Europe, began with a fleet reported at ten electric SUVs — Arcfox Alpha T5 vehicles from the Chinese group BAIC — with trained safety operators on board during the early phase, rides priced at just under two euros, daily operation from 07:00 to 21:00, and a service area of roughly 90 square kilometres covering the wider centre of Zagreb and the airport. The step from that to two thousand cars is a factor of two hundred, and it is announced rather than delivered. There is also a contrast worth stating plainly. In the same month that the United States closed its door to new Chinese robots and vehicles seeking fresh equipment authorisation, a Chinese autonomous driving company set out a four-city expansion in Europe — arriving not as an importer, but as the technology supplier to a European fleet owner holding European licences. The regulatory question in Europe is not whether the software comes from China. It is who is legally answerable for the vehicle on the street, and in this deal the answer is a company from Zagreb.
Pony.ai Robotaxi (Gen-7) →Figure builds its 1,000th Figure 03 — and paints it gold
Figure has manufactured its one-thousandth Figure 03 humanoid, chief executive Brett Adcock announced on 23 July 2026, marking the unit with a golden finish that observers immediately compared to C-3PO. The number matters less than the slope behind it. In late April 2026 the company said its BotQ plant had reached a throughput of one robot per hour, with just over 350 units built to that point. Crossing a thousand roughly three months later means the line held that rate rather than hitting it once for a press photo — which is the harder part of manufacturing, and the part most humanoid programmes have yet to demonstrate. For context on where the field sits: Tesla's dedicated Optimus line in Fremont was still not confirmed to be producing as of early August 2026, and most humanoid makers remain in pilot deployments rather than series output. For scale, this still leaves the American company behind the Chinese volume leaders rather than ahead of them: AgiBot reported some 15,000 cumulative units by June 2026, and Unitree shipped more than 5,500 machines during 2025. Figure has the steepest recent slope, not the largest number, and it has it on a single full-size model built on one line, for which the company states a target capacity of 50,000 units a year. None of this says how many of those thousand robots are doing paid work. Unit counts measure a factory, not a market, and the humanoid industry has so far been much better at the first than at the second.
Figure 03 →AgiBot A2 walks 106 km from Suzhou to Shanghai — and the fine print matters
An AgiBot A2 Ultra covered 106.286 km between 10 and 13 November 2025, taking the Guinness World Record for the longest journey by a humanoid robot. The robot navigated autonomously to defined waypoints and, according to AgiBot, was a standard production unit rather than a special build. The record is a battery-logistics result as much as a walking one: the distance was spread over several days, included a rest period, and the crew hot-swapped the packs, so this was not a non-stop march. That caveat is the interesting part — hot-swappable power, not longer endurance per charge, is what currently lets a humanoid stay on its feet for days.
AgiBot A2 →