ACQUIRING TARGETS...
· DroidHunt news
Mykhailo Fedorov — Ukraine's former defense minister and architect of the 2022 Army of Drones program — announced the "Army of Robots" on Sept 26 at the IT Arena conference in Lviv and on X. The private-sector "ecosystem" will invest in defense-tech companies, build its own projects and test with the military, with humanoid robots slated initially for mechanical front-line tasks: reloading machine guns, swapping batteries, placing sensors, casualty evacuation, mining and demining. The stated goals are aggressive: the first enemy soldier killed by a humanoid robot within 4–6 months, and the first infantry-less assault — robots and drones coordinating over shared protocols — within 12 months. None of this exists yet: the launch video is AI-generated concept renders, no hardware or organization behind the initiative was named, and Fedorov's team did not respond to press questions. He is recruiting engineers, a CTO and investors; Palantir CEO Alex Karp personally backed Fedorov's separate defense-tech company earlier this month.
Why it matters: Coming two days after Japan's Defense Ministry scoped non-combat humanoid roles, this is the second defense establishment in a week moving humanoids into military planning — and the first to state a combat-kill timeline. It is a declared ambition from a credible wartime tech organizer, not a capability; the 4–6 month mark (late March 2027) is the checkpoint to hold it against.
Sources: kyivindependent.com yahoo.com dev.ua
Sharpa unveiled three self-developed products at IROS in Pittsburgh on Sept 28 (booth 514). The D01 is billed as the first integrated tactile-sensing dexterous-manipulation humanoid: arm payload-to-weight ratio near 1:1, end-effector speed above 10.5 m/s, 1,000 Hz communication, 0.2 mm repeatability, and a full-body electronic skin sampling at 100 Hz across a 0.1–20 N force range with 0.2 N resolution — enough to catch a falling baton or react to being touched. The W02 dexterous hand packs 21 active degrees of freedom into a package roughly 30% smaller than its predecessor W01, with high-resolution visuotactile fingertip sensors (5 mN–30 N, 1 mm spatial resolution), palm e-skin, zero grasp radius for small-diameter objects like chopsticks, and IP54 sealing. The AE01 exoskeleton data glove closes the loop for teleoperation and training-data capture: 22 encoders map the operator's hand to the robot while each fingertip gets 256-level continuous haptic feedback; it is calibration-free and hardened against electromagnetic, light, humidity and temperature interference.
Why it matters: This is a full-stack tactile play — body, hand and data-capture device from one vendor — aimed squarely at the industry's current bottleneck, high-quality contact-rich manipulation data. Booth demos (baton catching, teleoperated bartending, toy catching) put it in direct competition with the dexterous-hand and data-glove offerings multiplying across China and the US.
Sources: leiphone.com
Tags: humanoid product-launch
Agility Robotics chief robot officer Jonathan Hurst confirmed to The Robot Report on Sept 25 that the company is "exploring a range of robot designs and form factors, including wheels," after a wheeled humanoid rendering appeared at the end of the Digit 5 launch video. Hurst said legs will address more applications as safety and reliability improve, but other form factors could bring Agility's technology into more environments sooner. No decision has been made on whether the concept becomes a wheeled Digit or a separate platform — Hurst's answer was "Too early to say!" — and there is no development or commercialization timeline. It is an early design exploration, not a product announcement.
Why it matters: The industry's flagship bipedal pure-play publicly hedging on form factor is a signal that the legs-vs-wheels argument is being settled market by market rather than ideologically. It also lands mid-SPAC: Agility's $2.5B Churchill XI merger is pending, with 2025 revenue of just $1.8M against heavy losses — investors will read "more environments sooner" as a revenue-timeline statement.
Sources: therobotreport.com humanoid.guide
Benmo Technology fixed its H-share offer at HK$21.60, raising about HK$1.08B from 50M shares with cornerstone investors taking 43.54%; CITIC Securities (HK) is sponsor. The stock lists on the HKEX main board on Sept 29 under code 6731.HK via the Chapter 18C specialist-technology regime — the first public listing of a pure direct-drive robot-component maker. Benmo's pitch is the reducer-free drivetrain: per Frost & Sullivan data in the prospectus, it held 61.1% of China's consumer-robot direct-drive power-module market by 2025 revenue, shipping about 8.5M units in 2025 and 5.7M in H1 2026, with four of the global top-10 consumer robot companies as customers (Roborock, Narwal, Dreame among the named ecosystem). Losses are narrowing on an adjusted basis: RMB 61.6M (2024) to 43.2M (2025) to 27.5M (H1 2026); the headline 2025 net loss of RMB 881M is mostly a RMB 770M non-cash redemption-liability revaluation.
Why it matters: After Unitree (STAR) and Mech-Mind (HKEX), this is the third embodied-AI supply-chain listing of the season and the cleanest test of public-market appetite for robot component plays. Founder Zhang Di says he wants the listing to set the first market price for the entire direct-drive category — debut-day trading will be read as a verdict on direct-drive versus reducer architectures for humanoid joints.
Sources: tmtpost.com inews.hket.com news.qq.com
Tags: supply-chain ipo
Researchers at Stanford's TML lab and Caltech released HomeBody, a system that lets a frontier vision-language model (GPT-6 Astra) directly orchestrate a humanoid's composable skill library — Pick, Place, Open drawer, Pick-from-drawer, Navigate — with no trained VLA policy layer in between. The G1 first explores an unfamiliar kitchen on its own, recording camera frames (RealSense D435i), LiDAR/SLAM geometry, joint poses and waypoints, then builds a geometrically and semantically faithful digital twin in NVIDIA Isaac Sim (Real2Sim) as spatial memory. That memory lets it resolve vague instructions ("clean up the kitchen," "fetch my medicine and toss the spoiled carton") by chaining skills, locating objects no longer in view, and self-correcting failed grasps — low-level execution is handled by segmentation (SAM 2.1), tracking memory (SAMURAI), stereo depth (Fast-FoundationStereo) and a pretrained AMO whole-body controller. Reported limitations are practical: VLM inference latency, overheating finger servos and high compute cost. The project page and videos are public; The Decoder reports the code is on GitHub.
Why it matters: HomeBody is a clean architectural data point in the VLA-versus-skill-library debate — a frontier generalist model coordinating reusable robot skills with zero environment-specific training. If the pattern holds up under independent replication, it lowers the training burden for long-horizon household tasks; the reliance on a predefined skill library is the obvious ceiling.
Sources: tml.stanford.edu the-decoder.com
Tags: research embodied-ai
AgiBot (Zhiyuan) and cookware brand Aishida kicked off a three-day, three-city store livestream on the morning of Sept 28, with the Lingxi X2 humanoid formally "onboarded" at Aishida's Wenling flagship — employee badge included — working as an in-store promoter from 10:00 to 20:00. A 100-city livestream plan sits behind the stunt. File it as deployment theater rather than industrial deployment: a real store, real shoppers and a real shift, but a marketing-led format.
Why it matters: Consumer-facing retail shifts are becoming a repeatable commercial pattern for Chinese humanoid makers (Qiyuan's store network, UBTECH's Leshan service project) as they hunt revenue beyond data-collection contracts.
Sources: stockdatasource.com qingyise.net
Tags: humanoid deployment
← Newer: Sep 29, 2026 Older: Sep 26, 2026 →
Home Every digest Robot catalog Drones on DronePriceHunt Privacy & affiliate disclosure RSS