Today's Highlights
· XPeng's robotics business raises over $900 million in its first funding round, at a post-money valuation above $6.3 billion
· China's Ministry of Industry and Information Technology (MIIT) seeks public comment on a humanoid robot standards framework, targeting at least 100 key standards by 2028
· Amazon's Zoox begins charging fares in Las Vegas, moving the first steering-wheel-free Robotaxi in the US into commercial operation
· Vision-driven soccer skills from a Tsinghua team led by Zhao Mingguo make the cover of Science Robotics's humanoid special issue
· Unitree Robotics (Chinese humanoid maker) closes down 10.31% on the 24th, market cap 243.9 billion yuan, a 44.8% drawdown over four trading days
· Mech-Mind opens its Hong Kong IPO order book, aiming to raise up to roughly HK$2.7 billion at the top of the range
Research Progress
Vision-driven reactive soccer skills let a humanoid play a full match using only onboard cameras · locomotion
Science Robotics's August humanoid special issue put on its cover a collaborative paper from Zhao Mingguo's team in Tsinghua University's Department of Automation, ByteDance Seed, and China Agricultural University — a rare cover placement for Chinese humanoid locomotion control research. Rather than chasing a single motion metric, the work fuses visual perception, state estimation, bipedal locomotion, and physical contests into one continuous task: using only onboard vision, the robot finds the ball, chases it, adjusts its gait, and shoots in multiple directions on a dynamic field. Methodologically, the team extends adversarial motion priors to perception-aware scenarios, using an encoder-decoder paired with a virtual perception system that simulates motion blur, lighting changes, and occlusion, so the policy can recover privileged state from incomplete observations. The hardware platform was provided by Booster Robotics (Chinese humanoid robotics maker), with validation conducted at a real RoboCup competition.
Yushi Wang et al. (Tsinghua University · ByteDance Seed · China Agricultural University) · Science Robotics 10.1126/scirobotics.aed1152 · Project page humanoid-kick.github.io
Q-Planning: freeze the behavior-cloning weights, let a small Q-function learn from failure · manipulation
Behavior cloning has a dead end: a policy that fails learns nothing from the failure unless it's fed more human demonstrations. This paper attaches a small off-policy Q-function to a billion-parameter-scale vision-motor policy, using it at inference time to value-weight multiple BC samples; online self-improvement fine-tunes only the Q-function while leaving the BC weights untouched, sidestepping the scalability problems of RL fine-tuning for large models. On two contact-rich real-robot bimanual tasks, five rounds of iteration on rollouts generated purely from the model's own deployment raised cup-stacking from 40% to 90% and wallet insertion from 25% to 80%; under the same budget, SFT on successful rollouts alone plateaued at 55% and 30% respectively. On LIBERO-10, results went from 93% to 99%; on RoboTwin, from 83.8% to 91.4%.
Varun Giridhar et al. (Georgia Tech) · arXiv 2608.21204 source
ForeTime-VLA: distills a world model into four tokens to catch moving targets on a conveyor belt · vla
Grasping moving objects requires a policy to anticipate the moment of contact, but VLAs typically only fine-tune on the current frame, while running a video-scale world model teacher live at deployment is too costly. The authors compress a frozen teacher's future video latents into a 64-dimensional target, using only an eight-frame history encoder online to predict it, alongside four future tokens and a phase token fed into the VLM prefix, keeping inference causal. Across three conveyor-belt speed settings on a real robot, the method completed 44/90 grasps versus 23/90 for π0.5; at the fastest setting, 11/30 versus 2/30. Success rate on stationary targets was 81.1%, on slow-moving targets 58.9% — 12.2 and 22.2 percentage points above the next-best baseline, respectively.
Siyuan Ma et al. · arXiv 2608.20735 source
WA-JEPA turns V-JEPA into a driving-capable world-action model · autonomy
V-JEPA learns representations via random masked completion and deterministic regression, a paradigm poorly suited to driving planning, which needs to be both future-oriented and tightly coupled with action. The changes are threefold: random spatiotemporal masking is replaced with hybrid future-masking pretraining, deterministic regression is replaced with conditional flow matching in latent space, and a joint predictor denoises future scene tokens and the ego-vehicle trajectory together in the same latent space, letting action supervision directly shape the world representation. After pretraining on nuPlan and fine-tuning on NAVSIM, the model scores 91.7 EPDMS on NAVSIM-v2, beating the strongest end-to-end and world-action baselines by 1.6 and 1.3 points; on closed-loop HUGSIM, without dedicated fine-tuning it achieves the best HD-Score of 0.4462. Code has been released.
Xinlin Wang et al. · arXiv 2608.20974 source
GhostTac: faking a robot's sense of touch with electromagnetic interference, no contact needed · perception
The physical-layer security of tactile sensors has barely been examined. This paper presents the first contactless attack, exploiting the nonlinear rectification and limited bandwidth amplification of front-end circuits to turn a crafted electromagnetic interference signal into a sustained DC bias that bypasses onboard filtering, stably corrupting readings — and enabling fine-grained manipulation by tuning the interference's spatial distribution and amplitude. The authors reproduce the attack successfully across 10 sensing modules, two dexterous hands, and 15 tactile sensor models, and demonstrate it derailing three case studies — grasping, slip detection, and material classification — with consequences including robots applying excessive force that damages objects or injures people.
Kun Wang et al. · arXiv 2608.20817 source
Logic-VLA adds a formal-constraint channel to VLA · vla
Natural-language instructions can't precisely express safety boundaries or temporal requirements — something like "don't fly into that zone for more than three seconds" is hard for a policy to execute exactly. Logic-VLA additionally accepts a Signal Temporal Logic specification at inference time, pretraining a grammar-graph encoder to capture temporal-logic semantics, then adapting in two stages: supervised fine-tuning on demonstrations that satisfy the specification, followed by preference optimization on paired satisfying/violating trajectories. In closed-loop quadrotor navigation tests, on STL formulas unseen during training, the specification-satisfaction rate is 24.8 to 40.7 percentage points higher than a specification-free baseline, with at most a 1.8-point loss in original task success rate. One policy handles multiple constraint types, rather than requiring a separate model trained per specification.
Celina Shiyu Wang et al. (USC) · arXiv 2608.20556 source
Koala Gripper: co-designing the hand that collects data and the hand that does the work · manipulation
Handheld data-collection devices are typically modeled after existing parallel-jaw grippers, which are awkward for humans to use and degrade the quality of the manipulation data collected. This design from RAI Institute builds the constraints of both data collection and execution into a single co-design process, yielding a mechanically optimized finger/trigger linkage, an integrated dual-thumb structure, and a grip that better matches the human hand. On the robot side, the fingers are back-drivable with an equivalent mass of only tens of grams, enabling stable grasps across a wide range of object sizes, forceful tool use, and fine single-item separation. The team also validated the design end-to-end with a full data-collection-plus-policy-execution pipeline.
Amar Hajj-Ahmad et al. (RAI Institute) · arXiv 2608.20546 source · Project page koalagripper.rai-inst.com
The "right to be forgotten" for demonstration data — existing metrics can't audit it · benchmark
When someone later demands the removal of robot demonstrations they contributed, full retraining is the standard answer, but its cost scales with policy and dataset size. The authors argue that forgetting losses or single-shot membership-inference attacks borrowed from machine unlearning can't tell you what an edit actually removed from a closed-loop policy, and propose a dual-axis audit calibrated against retraining as the reference: a behavioral axis that measures how much the edited policy's actions diverge from a retrained version on matched states, with the lower bound set by variance across multiple independent retrainings; and an evidence axis that runs membership attacks per demonstration, reporting both ranking and absolute membership loss. Across five pre-registered experiments spanning three real-robot policy classes and two simulators, the two axes diverge in opposite directions — an edit can fix task performance while leaving just as much trace behind. On an ACT robotic arm, a redirection edit restored blind-evaluation success rate to 18 out of 20 trials.
Jiazhuo Li et al. · arXiv 2608.20784 source
Other papers today: VT-MUSE unifies vision and touch into a sequential representation instead of encoding and fusing them separately, beating the strongest baseline on all tasks (arXiv 2608.21290 source); CertVLA delivers a certifiable defense for closed-loop VLA control against bounded patch and texture attacks (arXiv 2608.20791 source); PhysCaP has code-as-policy agents actively explore through interaction, estimating object mass and stiffness from proprioception without training (arXiv 2608.21031 source); AURORA proposes a roadside-cooperative end-to-end driving framework with a companion V2X language benchmark (arXiv 2608.21032 source); ViTacPhys estimates mass, friction-coefficient class, and continuous stiffness from human visuo-tactile demonstrations and adjusts grasping accordingly (arXiv 2608.21355 source); GraphOp-WM uses a graph-operator world model to generalize to unseen combinations of link length, mass, and damping (arXiv 2608.20936 source); Action-JND introduces just-noticeable-difference modeling into VLA token compression, defining "noticeable" in terms of closed-loop action response (arXiv 2608.21247 source); TaPeR recovers a sparse task-priority graph from a small number of demonstrations, letting robots flexibly reorder subtasks (arXiv 2608.21035 source); one study uses human demonstrations recorded on hard ground to teach a 29-DoF Unitree G1 to stand up on deformable soft ground (arXiv 2608.20852 source); VisTa3D releases the first thin-object reconstruction dataset with synchronized RGB, depth, and tactile response maps (arXiv 2608.20740 source).
Funding & Deals
XPeng Robotics | First round | Over $900 million | Post-money over $6.3 billion · humanoid
IDG Capital led the round, with Loyal Valley Capital (Chinese venture firm) participating; Tencent and Alibaba joined as strategic investors, while XPeng Group retains control. The post-money valuation of roughly 43 billion yuan sets a new record for a single private-equity funding round in China's embodied-AI sector. IDG has a long history with this line — it also led the over-$100 million Series A for XPeng Robotics (formerly Peng Xing Intelligent) in July 2022. The use of proceeds is specific: software and hardware R&D, physical AI model training and iteration, high-quality data collection, building out a full-chain mass-production base, and global commercialization. The next-generation IRON humanoid is slated to enter mass production by the end of 2026, targeting monthly capacity in the thousands, equipped with a second-generation VLA model and three Turing AI chips; the mass-production base in Guangzhou's Guangtang Science and Innovation City covers roughly 110,000 square meters. The rollout plan starts with guided tours and retail assistance in XPeng stores and campuses, with formal launch and delivery in China and overseas targeted for 2027.Source: Kuai Technology source
Mech-Mind | Hong Kong IPO order book opens | Up to roughly HK$2.7 billion · industrial
Following approval of its listing hearing in mid-August, the company opened subscriptions on August 24, offering 23,140,590 H shares at a guidance price of HK$95.3 to HK$101.7, with subscriptions closing at noon on August 27 and listing on the Main Board on September 1 under ticker 09615. At the top of the price range, with the over-allotment option fully exercised, gross proceeds would be roughly HK$2.7 billion (US$344.4 million). Nine cornerstone investors have together committed US$186 million: Baillie Gifford $60 million, Taikang Life $40 million, Invus, Jane Street, Ghisallo, and Vision Knight Capital each $15 million, NGS Super Fund and E Fund Management each $10 million, and BYD's indirect wholly-owned subsidiary Golden Link Worldwide $6 million. Founded in 2016 by Shao Tianlan, the company doesn't build complete robots — it sells standardized hardware and software components such as Mech-Eye industrial 3D cameras and the Mech-GPT multimodal model, designed to be mounted on various robot platforms; its pre-IPO shareholder list includes Hillhouse (HSG), Qiming Venture Partners, and Intel.Source: KrASIA source
NEURA Robotics acquires ADLATUS Robotics | 100% stake | Amount undisclosed · embodied
The target company, based in Ulm, Germany, makes autonomous cleaning and sweeping systems for industrial, logistics, medical, commercial, and public-space use, with hundreds of units already deployed in the field and its own onboard navigation software. This follows NEURA's earlier addition of the Bosch Rexroth ACTIVE Shuttle to its lineup, both feeding into the same platform, Neuraverse, which lets robots of different form factors and from different manufacturers share AI models and sensor information. Founder and CEO David Reger put it bluntly: "We are not buying ADLATUS to add another cleaning robot to our lineup. We want to give this class of machines a new brain." The goal is to move cleaning robots beyond preset routes and task instructions, letting them identify floor material and dirt type themselves and decide on a strategy. NEURA recently closed a €1.2 billion funding round.Source: Brand Icon Image source
Weifan Intelligent | Seed+ round | Over 100 million yuan · hardware
Existing investors Yanchuang Group and Haiyi Investment increased their stakes, joined by a fund under CICC Capital, Jiukun Venture Capital, Zeyu Capital, Shanhan Chaoxin, Daohe Yuanqi, and Jiuzhao Capital; humanoid robot maker Songyan Dongli (Chinese humanoid startup) is also among the investors. This round comes roughly three months after the company's several-hundred-million-yuan seed round in May. Founded in May 2025 as a spinoff from Peking University's brain-inspired chip laboratory, the company integrates a "brain" handling perception and reasoning with a "cerebellum" handling motion control onto a single chip; the PAICORE 2.5 test chip, developed jointly with Peking University, has been taped out and returned, with its companion on-device deployment tool, OmniRT, set to open to a first group of partners in September. In an internal demo provided by the company, a GR00T N1.6 model optimized with OmniRT ran stably above 10FPS on a Jetson Orin AGX and above 20FPS on a Jetson Thor; for reference, the company cites a public benchmark in which the 3.3B-parameter π0 runs at 5–8Hz on a Jetson AGX Orin.⚠️ Vendor-reported figuresSource: TMTPost source
Embedd (London) | Pre-Seed | €2.3 million · adjacent
Seedcamp led the round, with Cocoa, Connect Ventures, 2100 Ventures, and others participating. The three Ukrainian founders' previous hardware company was hit first by pandemic-era chip shortages, then by the war, repeatedly forcing them to rewrite drivers for swapped-in components — so they turned that grind into a product: building digital twins of chips so an AI agent can use that context to auto-generate integration code, reading the thousands of pages of manuals so engineers don't have to. The company says it has helped customers speed up production-grade software delivery for chips by up to 6x, and after commercializing in April has signed several semiconductor customers, including providing Zephyr support for Microchip Technology.Source: EU-Startups source
Qingcheng Intelligent (Qingdao) | First round | Amount undisclosed · hardware
Shenzhen Investment Holdings Donghai led the round, joined by Qingdao Dinghui Investment, a platform under the Qingdao Shinan District government, along with supply-chain participants Hangzhou Tiankuan Technology, Yuanqi Innovation (Xiamen) Robotics, and Shandong Nuojin. Founded in 2020, the company focuses specifically on robot controllers, making its "Cloud Elf" controller and matching complete units. The position of Chinese-made controllers remains difficult — ABB and Yaskawa still dominate through deep integration between their control systems and robot bodies, and Chinese manufacturers still lag in high-end control algorithms and multi-axis coordination stability. On August 7, Qingdao Silicon-Based Intelligent Robotics Company — led by municipal- and district-level state capital, with Qingcheng Intelligent holding a 15% stake — was registered.Source: Guanhai News source
Commercialization & Deployment
Zoox begins charging fares in Las Vegas, moving the first steering-wheel-free Robotaxi in the US into commercial operation · autonomy
No steering wheel, no dashboard, no brake pedal — a four-seat vehicle that's fully symmetrical front-to-back and doesn't need to turn around in a dead end is now charging the public to ride. According to an August 23 Wall Street Journal report, Amazon's Zoox has launched paid service in Las Vegas, while its San Francisco operation remains a free pilot for select users. What opened this door was a temporary exemption approved by NHTSA on July 30 this year: current federal motor vehicle safety standards assume a human driver is present in the vehicle and impose hard requirements for steering wheels, brake pedals, mirrors, and wipers — requirements Zoox's design inherently lacks in part. The exemption runs for two years, permitting deployment of up to 2,500 exempted vehicles per 12-month period for commercial operation; the vehicles are barred from roads with speed limits above 45 mph, and excluded from operating in heavy rain, snow, standing water, or heavy leaf cover. NHTSA's review covered only the vehicle safety performance tied to the exemption, not an overall safety assessment of the self-driving system itself. Before charging fares, Zoox's dedicated fleet had logged over 3 million driverless miles on public roads and carried nearly 1 million riders across four cities. Scale remains a weak point — the current fleet is about 100 vehicles, versus roughly 4,000 for Waymo; its Hayward factory covers 220,000 square feet, with a long-term annual capacity target of up to 10,000 vehicles, and the production-intent version unveiled in June is capable of a rate of 100 vehicles per week. Co-founder and CTO Jesse Levinson recently told Axios that scaling the technology will require investment in the tens of billions of dollars.Source: Guancha source
UBTECH (Chinese humanoid robot maker) talks repeat orders, not unit counts — U1 series tops 13,000 orders across all channels · humanoid
Speaking at WRC, UBTECH Vice President and head of its Embodied AI and Humanoid Robotics Research Institute, Jiao Jichao, said the company cares more about repeat purchases from industrial customers than shipment volume; in 2025, UBTECH sold 1,079 humanoid robots for 820 million yuan in revenue, 41.1% of total revenue, making it the company's largest revenue source. The U1 series, launched in late June and priced from 119,800 to 990,000 yuan, has surpassed 13,000 orders across all channels, with the first batch of deliveries starting September 16.⚠️ Vendor-reported figuresSource: China Business Journal source
XPeng Robotaxi completes 2,000 internal-test trips in Guangzhou, targets driver-free rides next year · autonomy
On its Q2 earnings call, He Xiaopeng said its factory-built, mass-production Robotaxi equipped with second-generation VLA has completed over 2,000 internal-test trips in Guangzhou, validating the full passenger-carrying demonstration operation process, with cloud-based remote takeover platform development also complete. "The goal is to achieve driver-free passenger operation next year," with plans to partner with ride-hailing platforms in China and abroad in 2027 to expand into core cities, monetizing through vehicle sales, technical services, and operational revenue-sharing. He also disclosed that the company recently formed a group-level business-development team, currently in talks with partners in China and overseas to license out its Turing AI chip, second-generation VLA, Robotaxi, and humanoid robot technology. Between internal-test volume and removing the safety operator lies regulatory approval.⚠️ Forward-looking, company guidanceSource: LatePost source
Mercadona opens its first semi-automated warehouse in Madrid, 70 robots handling 500 orders a day · industrial
Spanish supermarket chain Mercadona's seventh online-fulfillment warehouse, located in Villa de Vallecas, Madrid, covers 32,000 square meters, cost €54 million to build, and is the largest in its network — as well as the first to introduce partial automation. Seventy robots manage over 2,700 dry-goods SKUs, delivering items directly to workstations and eliminating the need for pickers to walk the aisles. It currently handles over 500 orders a day, with a designed capacity of 5,000; combined with its Getafe and Boadilla del Monte warehouses, daily processing capacity across the Madrid region rises to 8,000 orders. The center provides 700 jobs.Source: Modernet Digital source
AgiBot (Chinese embodied-AI startup) signs deal with Chimelong to build an embodied-AI theme park · embodied
The two companies signed a strategic cooperation agreement at Chimelong Spaceship Park in Hengqin, with plans including an immersive robot-themed park, a "Robot Animal Kingdom," a large-scale robot circus show, a cyber-themed parade, and a robot-service hotel, along with a joint embodied-AI lab focused on cultural-tourism applications to be based at Chimelong. AgiBot claims to already be the world's largest embodied-AI company by both shipment volume and revenue. The agreement has not yet disclosed which robot models, unit counts, or opening date will be involved.⚠️ Forward-looking, company guidanceSource: iFeng Tech source
Industry Developments
MIIT seeks comment on national humanoid robot standards framework, targeting at least 100 key standards by 2028 · humanoid
The draft "Guidelines for Building China's National Humanoid Robot Industry Standards System (2026 Edition)" is open for public comment from August 25 to September 23. The target is stated plainly: by 2028, complete at least 100 key standards covering humanoid robot capability testing and evaluation methods, key technologies, platforms and systems, application scenarios, and safety governance, with over 200 companies involved in promoting and implementing the standards. A few provisions in the framework are worth flagging early: it standardizes a capability-tiering methodology for humanoid robots, standardizes the definition, coding rules, and identification mechanism for unique IDs, sets requirements for operational capability and environmental adaptability geared toward structured and semi-structured industrial environments, and lays out ethical constraints and social-impact assessment requirements for technology development and application. The document notes that over a hundred humanoid robot standards — in China and elsewhere — are already in development or pre-development, with inconsistent terminology, interfaces, and testing/evaluation criteria being one of the root causes behind why "everyone's numbers look good" in the industry today.Source: Xinhua source
Unitree closes down 10.31%, market cap 243.9 billion yuan · humanoid
On its fourth trading day since listing, shares closed at 603.08 yuan on August 24, down 10.31%, with total market cap falling to 243.9 billion yuan — now second on the global humanoid-robot market-cap ranking. Against the first day's intraday high of 1,100 yuan, the stock has fallen 44.8% cumulatively, corresponding to a drop in market cap from 444.9 billion yuan, erasing over 200 billion yuan. The prior three trading days saw declines of 18.70% and 2.12%, followed by today's drop. The IPO price was 150.80 yuan, with an offering P/E of 219x versus an industry average of 38.56x. The company shipped over 5,500 humanoid robots in 2025 (excluding wheeled dual-arm models), ranking first globally by shipment volume.Source: Beijing Daily source
China's embodied-AI funding hits 93.5 billion yuan in H1, WRC booths shift from demos to POCs · adjacent
According to IT Juzi, total funding in China's embodied-AI sector reached 93.5 billion yuan in the first half of 2026, up fivefold year-over-year. The quality of exhibits at trade shows is shifting too — Wujie Dongli (Chinese robotics startup) brought its "human-robot symbiosis" space, built in collaboration with South Korea's HOLLYS Coffee, to the show floor; Galbot (Chinese robotics startup) built a 1:1 replica of its front-warehouse collaboration with JD.com; and Zibianliang (Chinese robotics startup) had robots fold clothes and scoop cat litter live. Xingyuan Zhichuang (Chinese robotics startup) founder Liu Dong estimated total industry shipments this year at roughly 70,000 units and argued that interactive world models can decouple the robot body from the model, letting a small amount of real-robot data map a model onto different robot platforms. Robot Era (Chinese humanoid startup) co-founder Xi Yue offered a more grounded read: in 2025, automakers and large logistics firms ran POCs and collected data with little regard for cost, but in 2026 customers are starting to seriously calculate ROI, especially in logistics; adapting to a new scenario used to take two months, now it can be done within a week. Guanghui Intelligent (Chinese data company) founder Yang Haibo said customer data demand is a hundred to a thousand times what it was last year, jumping from hundreds or thousands of hours to over a hundred thousand or even a million hours. By incomplete count, over 20 embodied-AI companies have been reported this year to have IPO plans or other capitalization arrangements underway.Source: China Entrepreneur source
Beijing Humanoid Robot Innovation Center opens Pelican-VL 2.0 for commercial use; Tien Kung Omni launches in three product lines · world-model
During WRC, the Beijing Humanoid Robot Innovation Center made its embodied-AI foundation model, Pelican-VL 2.0, fully available for public/commercial use, with reinforcement-learning post-training boosting its agentic capability, focused especially on long-horizon tasks and tool use; the company says it substantially outperforms GPT5.5 on embodied-AI benchmarks, though the benchmark names and methodology weren't disclosed. Also unveiled at the same event, Tien Kung Omni is positioned as a "Humanoid Native" open foundation platform, fully opening up low-level interfaces, motion-control frameworks, and embodied manipulation capability, and launching in flagship, standard, and basic product lines covering algorithm development, data collection, hands-on training, guided-tour interaction, and performance/demonstration use. The center also signed agreements with partners including Eutectical and Blue Holding, covering six countries and regions including China, Germany, South Korea, Japan, and Spain.⚠️ Vendor-reported figuresSource: IPO Zaozhidao source
ShengShu Technology (Chinese generative-AI startup) proposes a five-level roadmap for general world models; Motubrain adapts to a new robot body with just 50 to 100 demonstrations · world-model
Speaking at a WRC breakout session, Zhu Jun broke general world models into five levels: L1 world generation, L2 interacting with the world, L3 acting in the world, L4 autonomous world agents, and L5 world organizer, saying his team has already covered the first three, corresponding to its Vidu, Vidu S1, and Motus/Motubrain product generations. On the technical side, he emphasized a data pyramid narrowing from massive amounts of web video down to real robot interaction data, and treats failed attempts and correction processes as valid learning signal; architecturally, the system uses Mixture-of-Transformers (MoT) to give different modalities dedicated parameters, with shared attention handling cross-modal interaction. Motubrain, released this past April, unifies environment understanding, state prediction, and action generation into a single model, running roughly 10x faster than Motus at inference and adapting to a new robot body with only 50 to 100 human demonstrations; it reportedly scores 96.1 on RoboTwin 2.0, currently the top result, and has been validated across nearly ten robot platforms including Galbot, Star Dust Intelligence (Chinese robotics startup), and Qianxun (Chinese robotics startup).⚠️ Conference presentation, self-reported figuresSource: LeiPhone source
Xiaomi unveils three Xring chips; D100 is China's first 3nm autonomous-driving chip · hardware
Xiaomi unveiled three self-developed chips — Xring O3, O100, and D100 — the same day, all having completed full validation, spanning smartphones, AI compute, and autonomous driving. The O3 will debut in the Xiaomi 18 Fold, launching in September, while O100 and D100 will go into commercial use next year. The D100 uses a 3nm process, pairing a 20-core CPU with a 16-core NPU, supporting up to 160GB of unified memory, enabling on-device deployment of 200B-parameter models. Xiaomi founder Lei Jun subsequently posted a photo of Xiaomi's humanoid robot holding the three chips. The robot, nicknamed "Tie Da," made its public debut at this year's WRC, standing about 1.70 meters tall, weighing about 66 kilograms, with 66 degrees of freedom across its body, half of them concentrated in the hands.
Source: Kuai Technology source
XPeng posts Q2 revenue of 19.74 billion yuan, delivers 103,000 vehicles · autonomy
Total Q2 deliveries were 103,295 vehicles, up 64.8% quarter-over-quarter; total revenue was 19.74 billion yuan, up 51.5% quarter-over-quarter and 8.0% year-over-year; overall gross margin was 20.7%, with automotive gross margin at 12.1%; services and other revenue reached 2.70 billion yuan, up 93.9% year-over-year, driven mainly by technical R&D services provided to other automakers. Cash reserves stood at 40.48 billion yuan as of June 30. The Wall Street Journal covered the same earnings report the same day, headlining an expanding net loss. Its second-generation VLA recently completed local acceptance testing in Germany, with the company saying the same model can cover both Chinese and European road conditions, targeting delivery of localized driver-assistance systems to global customers starting in 2027. Q3 guidance calls for deliveries of 115,000 to 121,000 vehicles and revenue of 21.7 to 23.4 billion yuan.Source: Autohome source
Hardware & Supply Chain
· Dexterous-hand prices fall sharply: high-DoF dexterous hands that used to sell for 400,000–500,000 yuan have dropped to the ten-thousand-yuan range at this year's WRC — Yuequan Bionics (Chinese robotics maker) launched a product in that price tier, and Xinuo Future (Chinese robotics maker)'s high-DoF integrated arm-and-hand product is priced at 128,000 yuan; 36Kr reports China's dexterous-hand sector raised about 16.877 billion yuan in 2025, already surpassing 25 billion yuan in H1 2026 source
· Landpoint Technology (Chinese force-sensor maker) capacity plans: ⚠️ vendor-reported figures cited by 36Kr from GGII data, claiming a 72.6% share of the Chinese humanoid-robot six-axis force sensor market; the company says it has built automated production lines with a designed annual capacity of 1 million joint force-sensor units and 200,000 end-effector six-axis force sensor units, and has cut disclosed delivery times from the 6–8 weeks typical of some overseas products to 2–3 weeks source
· Miniature planetary roller screws remain a weak link: the smaller the size, the harder it is to control precision, machining, and batch-to-batch consistency — this is the toughest barrier between dexterous hands reaching ten-thousand-yuan pricing and achieving consistency at ten-thousand-unit production scale source
· Xinuo Future Prima 1: debuted globally at WRC as a fully direct-drive dexterous hand; the company says it can already reproduce about 90% of common human hand movements from a hardware standpoint, but complex high-precision tasks still rely heavily on teleoperation, kinesthetic teaching, or trained policy models
· UBTECH's two chip tracks: co-developing a dedicated on-device chip with MetaX (Chinese GPU maker) for software-hardware co-acceleration; separately partnering with BASiC Semiconductor (Chinese power-semiconductor maker) on humanoid robot battery-life improvements source source
Top comments (0)