ACQUIRING TARGETS...
· DroidHunt news
XPENG held its first Robotics Supply Chain Partner Conference on Sept 22, bringing together suppliers of joint modules, actuators and precision structural parts. In its Sept 23 announcement, the company said its robotics business has signed supplier designation agreements and is scheduling production on its manufacturing lines; it did not name suppliers, disclose contract values, or specify committed component volumes. The move follows the Sept 8 commissioning of XPENG's automated IRON production line and keeps the stated timeline intact: mass production by end-2026, initial deployment in XPENG stores and campuses, market launch and deliveries in China and overseas in 2027. CEO He Xiaopeng told suppliers humanoid robotics is more complex than automotive and requires longer-term co-investment; IRON also demoed orientation/following, multilingual conversation and identity memory at the event.
Why it matters: Supplier development is becoming part of the product itself. XPENG's conference is the corporate-announced counterpart to Tesla's reported Ningbo supplier audits and Apptronik CEO Jeff Cardenas's warning this week about US parts shortages — the race is now as much about securing consistent motors, gears and structures at scale as about the robots themselves.
Sources: prnewswire.com humanoidsdaily.com
Tags: humanoid production supply-chain xpeng
Feather Robotics formally launched on Sept 24 with $7.6 million in seed funding led by Gradient (Google's AI-focused fund), with Builder Capital, Geometry, Seed Innovation and Virgo VC participating. Its product: a wheeled, dual-arm humanoid starting at a promotional $29,990 (list $32,990), aimed at developers who integrate their own AI models via Feather's controls and SDK. The design bets on simplicity: two 7-DoF arms, a torso with 600 mm of vertical travel, an omnidirectional base (1.3 m/s max), 1.8 kg rated payload per arm (6.8 kg peak), and 10-hour battery life with dual hot-swappable packs. The stack supports Python and ROS 2, built-in teleoperation (including Apple Vision Pro and Meta Quest), simulation, collision detection, interchangeable end effectors and swappable onboard compute. CEO Hoa Mai co-founded Kind Humanoid (acquired by 1X in January 2025); co-founder/CPO Parsa Bakhtiari scaled Tesla Model 3 production. Manufacturing is underway in the Bay Area at up to 30 robots/month, with a next facility planned for 200/month; the company cites roughly a year of customer field testing but has not published reliability or deployment figures.
Why it matters: A credible, low-cost mobile-manipulation platform with fresh capital lowers the entry price for applied humanoid development — and tests whether "hardware company, third-party intelligence" is a viable wedge against vertically integrated players like Figure and Tesla.
Sources: humanoidsdaily.com dealroom.co
Skild AI published "Physical Self-Play" on Sept 23: its S1 robotics foundation model accumulated more than 140 years of simulated soccer experience inside NVIDIA Isaac Sim, competing against recent versions of itself with goal-scoring as the only objective. Dribbling past defenders, shielding, tackling and recovering from falls emerged without hand-crafted rewards. Skild then transferred the policy to a physical humanoid (a Unitree G1 in footage provided by NVIDIA and Skild) and reports it can play against humans and other robots, with early passing and coordination in four-agent games. Important nuance: the full pipeline is not demonstration-free — before self-play, S1 trained on human-referenced drills with skill-specific rewards; the model outputs joint angles 50 times per second. The "140 years" is accumulated simulated match time; elapsed wall-clock time, compute budget and parallel-match count were not disclosed. The work follows Skild's August introduction of S1 and its report earlier this month of crossing $100M ARR ten months after first commercial deployment.
Why it matters: Self-play is a credible route to robot capability beyond the ceiling of human demonstration data — the same recipe that produced superhuman game-playing AI, now applied to a generalist physical model. If it transfers to commercial tasks (Skild proposes collaborative manipulation and social navigation next), it reduces the field's dependence on expensive teleoperation data collection.
Sources: skild.ai humanoidsdaily.com x.com
Vesoma, a Munich-based humanoid startup (with a presence in Limassol, Cyprus), came out of stealth on Sept 23. The company is led by co-founders of Synapticon (servo-drive specialist) and Easybrain, with Martin Riedmiller — former DeepMind research director and a pioneer of reinforcement learning for robot control — heading AI. Its pitch: building "physical general intelligence" and the robot body together in Europe. The team previously built the Kyle prototype; its first Vesoma 1 units are now under construction. Details on funding, specifications, partners and timeline have not been disclosed.
Why it matters: Europe's humanoid bench — NEURA Robotics, Agile Robots, Hexagon — gains a new entrant whose AI leadership is unusually strong on reinforcement learning, the technique most believe separates demo robots from adaptive workers. Watch for funding and first-customer announcements.
Sources: humanoidsdaily.com linkedin.com x.com
Tags: humanoid europe new-entrant vesoma
Unitree announced the Dex5-S on Sept 21: a human-scale (1:1) five-finger dexterous hand with 22 degrees of freedom, priced from $6,500 before tax and shipping (¥39,900 in China). All 22 joints are motor-driven and backdrivable — an external force can move the joint rather than fighting a locked gearbox — with impact-torque protection at each joint. A Pro version adds tactile sensing; the Standard ships without. Specs from Chinese launch coverage: 620 g, 102×187×26 mm aluminum-alloy shell; thumb and pinky at 5 DoF each, index/middle/ring at 4; 22 motors with dual-encoder control; 1 kg rated / 2 kg peak payload; 200 W peak power; 15–65 V input; gigabit Ethernet, USB and high-speed 485 interfaces at up to 1,000 Hz control frequency. It succeeds the Dex5-1 (20 DoF, 16 active + 4 passive, optional 94-point tactile array). Unitree has not published fingertip force, cycle life, or delivery schedules.
Why it matters: A backdrivable 22-DoF hand at $6,500 undercuts research-grade hands from Shadow (quote-based) and SharpaWave by a wide margin — continuing Unitree's pattern of collapsing component price floors (G1 at $13,500) and pressuring Western hand makers just as manipulation becomes the field's bottleneck.
Sources: x.com alphasignal.ai humanoidsdaily.com
Barclays' research team projects that one million humanoids by 2030 could imply roughly 10,000 tons of demand for PEEK (polyether ether ketone), a high-performance engineering plastic prized for light weight and self-lubrication. In a Gasgoo interview (Sept 17, recirculated this week), Yu Fei, technical GM of Kingfa Tech's robotics unit — one of China's largest modified-plastics suppliers — argues the hype has significantly outpaced practical value. His case: PEEK's fatigue performance lags metal by at least two orders of magnitude, a hard constraint for high-frequency reciprocating robot joints; at up to ¥1,000/kg it raises cost without clear benefit in most current applications; and 80–90% of humanoid materials remain highly customized with no standards. Real progress exists — Huaxiang Qiyuan's modified-PEEK planetary reducer weighs 270 g (-70% vs metal) and cuts 5.3 kg per robot — and forecasts put humanoid-sector PEEK demand above ¥4B by 2027 (vs a ¥6.1B total PEEK market in 2024). Tesla, Unitree and AgiBot have begun incorporating PEEK parts.
Why it matters: A rare on-the-record materials reality check from inside the supply chain. PEEK concept stocks have repeatedly surged on humanoid enthusiasm; Kingfa's caution — from a company positioned to profit from the boom — is a useful corrective on how much of the materials story is booked revenue versus narrative.
Sources: humanoidsdaily.com autonews.gasgoo.com
Tags: humanoid supply-chain materials kingfa
← Newer: Sep 26, 2026 Older: Sep 24, 2026 →
Home Every digest Robot catalog Drones on DronePriceHunt Privacy & affiliate disclosure RSS