Humanoid factories, defence contracts and the robot stack

Differential Robotics completed a Series A1 of hundreds of millions of yuan, with returning investor Lightspeed China Partners making an additional investment. The financing was reported on March 31 and is earlier funding context for this briefing. It supports product development, recruitment, commercial applications and working capital.

The Hangzhou company develops autonomous flying robots and fleet-control systems. Its technology combines sensor fusion, navigation without satellite positioning, obstacle avoidance, exploration planning and three-dimensional reconstruction. These capabilities are intended for surveying and inspection in complex or hazardous spaces, reducing the need to send people into confined areas or rely on manually piloted drones.

Doozy Robotics announced expansion across the U.S., Gulf region and Asia, citing existing seed backing from investors including Cocoon Capital and a planned Series A. The company did not disclose a newly closed financing amount in the announcement.

Its industrial platform combines mobile robots, autonomous forklifts and a planned Industrial Super Humanoid under Eywa-OS, which allocates tasks across the fleet and responds to production changes. Customers subscribe to the integrated systems. Doozy reported paying customers on two continents, while the humanoid launch remained scheduled for Q3 2026. Its cited $144 million MoU and qualified pipeline were prospective commercial opportunities.

OriginFlow disclosed more than RMB500 million raised across angel, strategic and Pre-A1 rounds during its first five months of operations. The aggregate is not the size of the latest round. Monolith Capital led the Pre-A1; earlier financing included BlueRun Ventures, Oasis Capital and strategic backers including 58, Puhua Capital and the Tsinghua Seed Alumni Fund.

The company develops NeuroScale, a data-capture approach using surface electromyography to record muscle signals during human manipulation. Its PULSE model reconstructs hand poses, exerted force and tactile information, providing training data for robot skills that are difficult to learn from images alone. The financing supports this hardware, model and data infrastructure for dexterous manipulation.

Humanoid signed a binding phased deployment agreement targeting a four-digit number of wheeled robots across Schaeffler facilities by 2032. The first phase runs from December 2026 to June 2027 at Herzogenaurach and Schweinfurt, combining box handling with integration tests and validation of sustained operation.

The Robots-as-a-Service arrangement includes fleet software, maintenance, updates and support. A separate five-year supply agreement makes Schaeffler the preferred actuator supplier, covering more than half of Humanoid’s wheeled-platform demand through 2031. The deployment total remains a phased future commitment and should be kept distinct from annual factory capacity or units already manufactured.

Hexagon Robotics and Schaeffler expanded their partnership following a successful 2025 pilot, with Schaeffler planning to deploy at least 1,000 AEON humanoids across its global factories over seven years. The figure describes a planned rollout, rather than robots already installed. Schaeffler will also supply high-precision actuators for AEON.

AEON uses wheeled locomotion, sensor fusion and manipulation capabilities for industrial work. In the pilot it loaded, unloaded and inspected parts at a multi-machine station. The next stage extends deployment across sites, with further applications including automated part inspection targeted from the end of 2026. Production data gathered during use will support training for additional workflows and factory settings.

Zhejiang Humanoid and Jack Technology signed a programme for 2,000 humanoid robots customised for garment manufacturing. The agreement establishes a long-term supply and application-development relationship between the humanoid developer and the sewing-equipment company. The figure describes contracted systems, rather than 2,000 robots already working in factories.

The programme focuses on handling flexible fabrics, where varying material, wrinkles and alignment make conventional automation difficult. Zhejiang Humanoid reports that its NAVIAI systems can adapt trained manipulation strategies to different garment materials, including polo-shirt fabric and denim, with individual handling steps under ten seconds and stacked-piece alignment below two millimetres. Those company-reported demonstrations provide technical context for the order; the commercial programme still requires production and deployment of the customised systems.

Geekplus reported results from its Robot Arm Picking Station at Schneider Electric’s Shanghai warehouse, where it says the system doubled manual picking throughput and achieved at least 99.99% accuracy. It was production-ready within 48 hours, according to the company’s deployment account.

The station combines robotic manipulation with Geekplus Brain and existing mobile-robot infrastructure. Its zero-shot approach is intended to handle changing stock-keeping units without training a separate model for each item, while on-premise processing keeps warehouse data on site. The deployment adds automated item picking to the storage and transport functions already handled by Geekplus systems.

Romark Logistics selected DexoryView for inventory visibility at its Hazleton, Pennsylvania warehouse. Dexory’s autonomous scanning robot gathers inventory data during normal operations and feeds it into the software platform, supporting cycle counts without disrupting the existing shift structure.

The fully racked site handles high volumes of confectionery products. More frequent inventory measurements help staff locate discrepancies and resolve stock issues while keeping orders moving. The agreement also provides a pathway to additional Romark facilities and ongoing development of the warehouse workflow.

Primech AI signed a three-year lease for one Hytron restroom-cleaning robot at a high-traffic public-sector facility in Singapore, with deployment and recurring monthly revenue due to start no later than May 31. The agreement establishes a commercial reference site for the company’s cleaning system.

Hytron is designed to clean toilets, urinals, basins, floors and mirrors using brushes and electrolysed-water spray. The company says it can navigate between floors, use elevators, recharge and manage its water supply autonomously. Primech plans to extend leasing deployments in Singapore and pursue enterprise and distribution partnerships internationally.

Perennial Autonomy received a three-year, enterprise-wide indefinite-delivery/indefinite-quantity contract with a $500 million ceiling, coordinated by Joint Interagency Task Force 401. The ceiling sets the maximum value of procurement available under the agreement; it is not a booked order or realised revenue.

The procurement vehicle covers Merops interceptors, Bumblebee reconnaissance quadcopters and Hornet mid-range strike drones. Perennial combines computer vision, radio-frequency detection, jam-resistant communications and autonomous control to protect military personnel and facilities against aerial threats. The agreement gives U.S. military buyers a common route to order systems following operational deployments in Europe and the Middle East.

Latvia signed the first BLAZE interceptor procurement contract under a multi-year framework agreed with Origin Robotics in December 2025. The government approved the contract on April 21, and Origin announced it on May 19. The number of systems and contract value were not disclosed.

The framework provides continuity for future deliveries while allowing technical specifications to be renegotiated annually as the autonomous interceptor develops. Other European countries can join, and the agreement is eligible for EU SAFE financing. BLAZE is designed and manufactured in Latvia for counter-drone missions, giving the procurement programme a domestic production base while allowing buyers to receive updated systems over successive years.

Teledyne FLIR Defense launched FirstLook 125, a 2.6 kg throwable reconnaissance robot carrying visible-light and thermal cameras, illumination and two-way audio. The tracked platform is designed for looking inside buildings, rubble and confined spaces before personnel enter; articulated flippers help it negotiate obstacles and self-right after a throw or fall.

The company says the robot can withstand a five-metre drop. FirstLook 125 shares a controller and operating architecture with the Black Hornet 4 nano-drone, allowing one operator to switch between indoor ground reconnaissance and aerial observation without changing control equipment.

Notch launched its MAGIC software-defined antenna series for drones, ground stations, vehicles and fixed installations. Electronically reconfigurable metamaterials steer radio-frequency energy across 360 degrees without moving parts, allowing the antennas to work with existing radios and standard interfaces.

MAGIC WIZARD targets ground and fixed systems, weighing approximately 700 grams and drawing under 600 milliwatts. Notch claims more than twice the effective range of a conventional omnidirectional antenna in current configurations. The approximately 250-gram MAGIC ELF targets smaller aircraft with tighter payload limits. Initial configurations were available at launch, with wider frequency coverage and lighter versions planned, giving operators a way to improve communication links without replacing their radios.

HOMY Robotics presented an ageing-at-home companion service built on SoftBank Robotics’ Pepper hardware and care-focused software. The offer included a $200 refundable reservation and founding-family prices of $499 per month or $10,000 to own, with initial cohorts planned for late 2026.

The proposed functions include conversation, medication reminders, fall alerts, video calls and family updates. HOMY is adapting a mature social-robot platform to support an older person at home and help relatives check in remotely. Reservations and advertised capabilities describe an early-access offering; they do not establish that the planned home-care service was already deployed at scale.

EngineAI’s manufacturing programme combines its existing production line in Shenzhen’s Honghualing Industrial Zone with plans for a global manufacturing hub in Zhengzhou. The company described that expansion in January as preparation for production and delivery of PM01 and T800; the Zhengzhou facility should be treated as planned at this briefing’s date.

T800 is EngineAI’s full-size humanoid platform, offered in configurations for research, inspection, commercial services, logistics and industrial collaboration. Building manufacturing capacity around it requires tooling, component supply and repeatable assembly alongside the motion-control and manipulation software. The facility programme provides production context; it does not establish that a 10,000-unit line was operating or that a corresponding volume of robots had been delivered by May.

Matrix Robotics said its Matrix Fabrication Hall was heading toward first-year production capacity of 10,000 robots. The Shanghai factory programme supports the company’s MATRIX-3 humanoid, but the stated capacity target is not a count of completed deliveries. The manufacturing update accompanied its May AI Day presentation.

MATRIX-3 combines tactile sensing, a cable-driven dexterous hand and a motion-control model trained on human movement. Matrix positions the platform for manipulation in manufacturing, logistics and commercial services, with an early-access programme for industry partners. The fabrication hall is intended to turn that hardware and software programme into repeatable production; customer pilots and task validation remain distinct from the factory’s capacity ambitions.

Figure said its BotQ manufacturing facility had produced more than 350 Figure 03 humanoids and demonstrated a cycle time of one robot per hour, compared with one per day less than 120 days earlier. Dedicated module lines, supplier qualification and in-process inspections support the ramp; the company reported an end-of-line first-pass yield above 80%.

The growing fleet supplies robots for internal research, data collection, household-task development and commercial applications. Operating more units also gives Figure additional failure data for hardware revisions, diagnostics and service systems, while expanding the real-world experience used to develop its Helix robot models.

1X opened its 58,000-square-foot NEO factory in Hayward, California, with more than 200 staff and an initial annual capacity target of 10,000 robots. A planned San Carlos facility and further automation were intended to support more than 100,000 units annually by the end of 2027; these are capacity plans rather than delivered volumes.

The factory produces critical components including motors, batteries, structures, tendons and sensors, and handles assembly and testing. Early robots were being prioritised for internal development and home testing before consumer shipments planned for 2026. Keeping manufacturing and engineering together supports rapid design changes as NEO moves toward household use.

Tesla’s first-quarter update outlined plans to replace the Model S and Model X lines in Fremont with a first-generation Optimus production line designed for one million robots annually. Preparations for the large-scale factory were due to begin in the second quarter. Tesla also described a second-generation line at Gigafactory Texas with a long-term design target of ten million robots annually.

Optimus is Tesla’s humanoid programme for physical work, supported by its wider AI training and manufacturing infrastructure. Both robotics facilities were listed as under construction, and the company said first-generation production lines were being installed in anticipation of volume production. The figures describe planned capacity, rather than robots manufactured, orders received or a demonstrated production rate.

Leju Robotics and Dongfang Precision developed an automated humanoid production line in Foshan with annual design capacity in the 10,000-unit range. Dongfang’s account describes 24 assembly stages and 77 inspection and testing checks, with an AI control system coordinating production timing, materials and processes.

The partners stated that one robot could complete assembly every 30 minutes and that efficiency exceeded conventional manual assembly by more than 50%. The annual figure describes designed capacity, rather than confirming that 10,000 robots had been manufactured. Dongfang subsequently described the line during its May inauguration of the DF PRIMO robotics operation.

AGIBOT announced that its 10,000th robot had rolled off the production line. The milestone is a cumulative manufacturing figure across the company’s robot portfolio, rather than 10,000 deliveries of a single bipedal model or a count of independently verified autonomous deployments.

AGIBOT develops embodied-intelligence systems that combine locomotion, interaction and manipulation with physical robot hardware. The company attributed its production expansion to a more mature supply chain and standardised manufacturing, supporting the transition from small development fleets to larger commercial volumes. The manufacturing milestone was announced in March and is retained here as context for the broader production developments covered in the historical briefing.

The U.S. Navy approved Milestone C for Boeing’s MQ-25A Stingray after its first flight in April, clearing the programme for low-rate initial production. The decision authorises the Navy to award an initial production contract; it does not itself establish that production aircraft have been delivered.

MQ-25A is an uncrewed aerial refuelling aircraft designed to operate from aircraft carriers and extend the range of the carrier air wing. Its first flight autonomously followed a predetermined mission plan. Boeing’s engineering and manufacturing development programme includes four engineering development aircraft, five system demonstration aircraft and two ground test articles, providing the test fleet for the transition toward operational carrier use.

Hesai reported 118,282 robotics lidar shipments in the first quarter of 2026, up 137.8% from 49,731 a year earlier. Robotics accounted for approximately one quarter of its 471,723 total lidar shipments; the remainder were predominantly automotive driver-assistance sensors. These are sensor shipments, rather than a count of robots deployed.

The company’s lidar products provide three-dimensional perception for autonomous mobility and industrial, agricultural and service robots. Its results also described a 200,000-unit lidar design win with Zelos and an expanded supply relationship with Neolix. Those customer programmes illustrate the uses behind the shipment growth, while the design-win figure remains separate from sensors already shipped during the quarter.

XPENG announced that the first production unit of its GX-based robotaxi had rolled off the line in Guangzhou. The vehicle uses four in-house Turing chips, providing a claimed 3,000 TOPS of effective onboard computing, and a vision-based VLA 2.0 driving system without lidar or high-definition maps.

The production milestone follows a Guangzhou road-testing permit in January and creation of a dedicated robotaxi business unit in March. XPENG planned pilot operations in the second half of 2026 to test technical performance, passenger acceptance and the business model, targeting operations without an onboard safety officer by early 2027. The rollout therefore marks vehicle manufacturing readiness ahead of the planned service pilots, rather than a completed driverless commercial deployment.

Hugging Face researchers released LeRobot Humanoid, an experimental bipedal platform costing approximately $2,500 in parts, depending on sourcing, shipping and taxes. The release includes printable hardware, a bill of materials, wiring and assembly instructions, simulation assets and software for calibrating and controlling the physical robot.

Identification tools replay real robot data and fit simulator parameters, while MJLab environments support training and evaluating locomotion policies. The project connects building, simulation, data collection and real-world control in a reproducible research platform. Its initial hardware release focuses on the bipedal lower body, with upper-body integration and more advanced behaviours on the roadmap.

PsiBot released SynData, a multimodal dataset for imitation learning, manipulation and vision-action models. Its May 22 dataset card lists 449,363 clips across four subsets, including 313,674 egocentric clips and three glove-based subsets. The recordings combine natural human activity with structured measurements from PsiBot’s exoskeleton gloves.

Available channels vary by subset: head-camera RGB and depth, inertial measurements, wrist poses, hand joints, fingertip positions and tactile signals. The data is organised into indexed clips with task and language annotations, aligned on a 10-frame-per-second timeline and released under CC BY 4.0. This structure lets researchers pair what an operator sees with hand movement and contact information when training robot manipulation policies.

Researchers demonstrated a single reinforcement-learning policy for walking, running and fall recovery on the Unitree G1 humanoid. Their state-dependent adversarial motion prior routes training transitions between recovery and locomotion objectives according to body tilt, using three reference motion clips.

On physical hardware, a frozen policy runs at 50 Hz without explicit runtime commands to switch between modes. Experiments showed recovery from prone and supine falls and smooth transitions from walking to running under the same controller, providing a way to combine behaviours that are often managed by separate policies.

Researchers released RIO, an open-source Python framework connecting robot control, teleoperation, sensors, data formatting and learned-policy deployment. Its common interfaces let developers change robot hardware, grippers and cameras with less reconfiguration, reducing the integration work needed to reuse training and control workflows.

The study validated vision-language-action deployment on four hardware platforms spanning single-arm, bimanual and humanoid configurations. Demonstrations collected through RIO were used to fine-tune π0.5 and GR00T for household tasks including pick-and-place, folding and bowl scrubbing. The release provides infrastructure for running and studying robot models across physical setups.

ORBIT Robotics, an ETH Zurich Focus Project, revealed HELIOS, a four-armed prototype designed for microgravity work. Its arms provide both manipulation and anchoring: some can hold the robot in place while others handle equipment or cargo, allowing the system to work without legs in a space-station environment.

The tendon-driven architecture keeps motors near the shoulders to reduce moving mass, with cables transmitting torque to the joints and a rolling-contact elbow providing compliant motion. HELIOS follows the project’s IKARUS testbed for teleoperation, imitation learning and hardware iteration. The programme targets cargo, inventory and maintenance workflows that occupy astronaut time; the reveal establishes a research prototype rather than an operational robot already deployed in orbit.

Gecko Robotics is exploring Ouster’s Rev8 native-colour lidar as a way to add richer inspection data to its Cantilever operating platform. Gecko already uses Ouster lidar to navigate industrial environments and build digital twins; the announcement showed its Komodo robot equipped with a Rev8 OS0 sensor.

Rev8 produces structured three-dimensional measurements with native colour, ambient infrared and intensity information. Gecko intends to use those additional data layers to make structural anomalies easier to identify and interpret in its inspection software. The announcement describes sensor development and integration work, rather than quantified improvements from a completed customer rollout.

LG CNS and Kurly signed an MoU covering humanoid proof-of-concept testing and logistics automation at Kurly’s fulfilment centres. The companies will identify tasks where robots could reduce physical workload and improve safety, then measure task accuracy, execution speed and efficiency against existing workflows.

The tests will also evaluate PhysicalWorks, LG CNS’ robot learning and operations platform, using Kurly’s operating data. LG CNS plans to connect warehouse equipment and software across inbound handling, storage, picking and shipping. The agreement builds on its work at Kurly’s Gimpo and Changwon centres, which handle ambient, refrigerated and frozen goods, and establishes a validation programme before wider robot deployment.

The Office of the Under Secretary of War for Research and Engineering selected Shield AI to integrate Hivemind autonomy into the Low-Cost Uncrewed Combat Attack System, or LUCAS. The programme develops inexpensive one-way attack drones intended to operate in coordinated groups. An operational demonstration with one operator commanding multiple systems was scheduled for the autumn.

Hivemind will manage navigation, coordination and adaptation to changing conditions, while people retain control of strike decisions. The integration is intended to let a single operator supervise a group of aircraft in communications-constrained environments, extending Shield AI’s autonomy software into a new drone programme.

Kawasaki established the Kawasaki Physical AI Center San Jose, holding its opening ceremony on May 21. NVIDIA, Analog Devices, Microsoft and Fujitsu are initial collaborators. The centre will begin with healthcare and eldercare applications, combining AI with Kawasaki’s robotics and experience from operating manufacturing sites.

The programme includes autonomous service robot Nyokkey, indoor delivery robot FORRO, the hinotori surgical system and the CORLEO multi-legged vehicle. NVIDIA will contribute to AI and robotics development, Analog Devices to sensing and voice-enabled systems, Microsoft to cloud and AI infrastructure, and Fujitsu to healthcare systems integration. The San Jose centre will work with Kawasaki’s Japanese development sites and its Strasbourg innovation centre, which began operations in March.