Today's Highlights
· Unitree Robotics set its IPO price at RMB 150.80/share, implying an issue-date market cap of RMB 60.993 billion
· DeepSeek received a strategic placement allocation of over RMB 140 million in Unitree's IPO, with a 36-month lock-up
· Hadrian raised nearly $1.4 billion, valuation quadrupling in seven months to $7.9 billion
· LexisNexis's first humanoid patent report: China holds 73% of body-structure patents
· South Korea plans to deploy 1,000 AI robots per year, building purpose-built humanoids for 10 major industries
· GigaDevice (Chinese chipmaker) launched two robot MCUs in succession, with roughly 3 million units shipped for robotics applications in H1
· Nucleus emerged from stealth: deployed a humanoid robot into a German factory within 90 days
· DreamWAM lifted real-robot unseen-scenario success rate from 55.6% to 74.4%
I. Research Progress
DreamWAM: World Action Models Shouldn't Just Predict RGB · world-model
World Action Models (which learn action representations by predicting how the environment will evolve) typically predict the future in RGB pixel space, where task-relevant state changes get entangled with texture, lighting, background, and viewpoint noise. DreamWAM instead jointly trains four complementary future representations — appearance, motion, geometry, and semantics — while all these auxiliary branches are switched off at inference, keeping deployment purely RGB-based. The gains show up mainly out-of-distribution: under LIBERO-Plus perturbations, success rises from 51.36% to 63.44%, and on a real robot facing unseen lighting, backgrounds, and object placements, average success reaches 74.4%, versus 55.6% for the Fast-WAM-Joint baseline from the same group. Code and models are open-sourced.
Shanglin Yuan et al. (Huazhong University of Science and Technology, hustvl) · arXiv 2608.04996 source
MobileWAM: Moving World Action Models from Tabletops to Mobile Manipulation · world-model
Prior WAMs have been almost entirely confined to tabletop pick-and-place, while mobile manipulation demands simultaneous locomotion and whole-body manipulation amid scene-scale dynamic changes — something still largely handled with dynamics-free vision encoders plus hand-written coordination logic. MobileWAM uses a mixture-of-transformers to jointly attend a pretrained video diffusion transformer with a lightweight action expert layer by layer, while splitting the feed-forward layers into shared, locomotion, and manipulation experts that are softly routed by action intent. A training-time Chain-of-Foresight mechanism has intermediate representations predict future latent chunks progressively; at deployment the entire foresight chain and video generation are discarded, leaving only policy-level overhead. It outperforms existing mobile manipulation policies on ManiSkill-HAB and was fine-tuned and validated on a real ARX Lift2 robot.
Zehua Fan et al. · arXiv 2608.04657 source
BridgeVLA++: Giving 3D VLA a Spatiotemporal Memory · vla
Existing 3D VLA models are data-hungry, generalize poorly under distribution shift, and lack any explicit memory of past observations — making memory-dependent tasks (e.g., recalling where an object was placed after being occluded) largely infeasible. Built on top of BridgeVLA's original multi-view projection and heatmap intermediate representation, BridgeVLA++ adds a unified spatiotemporal memory module, achieving SOTA on two memory-dependent manipulation benchmarks without sacrificing the original version's data efficiency and generalization, and validated in bimanual settings and on an additional real-robot platform. Same-day HF trending score 7↑.
Peiyan Li et al. · arXiv 2608.05042 source
SiMDex: More Human Egocentric Video Isn't Always Better — Selection Matters · manipulation
Scaling human egocentric video for robot training is now conventional wisdom, but which portion of that data actually helps dexterous manipulation has remained unclear. SiMDex treats human-data selection for VLA post-training as a recommendation problem, using a retrieve-rank-rerank pipeline to select similar clips for each robot demonstration from a pool of roughly 32 million egocentric samples, operating entirely in an embodiment-agnostic action space without modifying the VLA architecture. Against a strong baseline using an equal amount of randomly sampled data, it uses only about 1.49 million clips (under 5% of the sample pool), yet raises overall success rate from 47.7% to 61.1%.
Nie Lin et al. · arXiv 2608.04196 source
Mind-VLA: Spatial Alignment Should Target the Object the Instruction Names · vla
Aligning representations with 3D scene geometry is a common way to improve VLA generalization, but such methods align the entire scene uniformly, overlooking exactly the geometry of the target object named by the language instruction — causing failures in fine manipulation and when the target is occluded. Mind-VLA first localizes the instruction-specified target, then aligns the VLA's latent representation with VAE and VGGT features from three views of that target. Using a compact 345M-parameter backbone, it achieves 93.9% on LIBERO and 4.47 on CALVIN; on real-robot occlusion tasks, average success reaches 54%, 32 percentage points above the best instruction-agnostic comparison method.
Xingyu Ding et al. · arXiv 2608.04633 source
Tactus: Open-Vocabulary Recognition Even from the Cheapest Pressure Arrays · perception
Tactile representation learning has for years centered on optical sensors that image gel deformation, while resistive pressure arrays — the cheapest and highest-volume sensor type — have been comparatively neglected. Tactus answers text queries using only pressure data, achieving a top-1 accuracy of 0.771±0.062 (four runs) on the STAG benchmark (27 object categories), matching and in the best case exceeding the 0.76 achieved by a supervised closed-set CNN on the same dataset, without training a classification head. The recipe is thoroughly small-data: 187 training recordings and 144,000 unlabeled same-sensor frames for masked-autoencoder pretraining, with the sensor's built-in calibration affine transform contributing more gain than all architectural changes combined. The paper also reports its failures — cross-sensor pretraining pooling yields no benefit, and joint training with vision actually hurts tactile performance; weights and code are open-sourced.
Abdul Basit Tonmoy et al. · arXiv 2608.04043 source
Deltoris: Real-Time Inference for Diffusion VLAs via Algorithm-Hardware Co-Design · vla
Diffusion-based VLAs deliver the best action quality and generalization, but at the cost of requiring 50–200 Hz control frequencies, imposing hard latency and energy constraints on edge devices. Deltoris exploits the temporal similarity between adjacent inputs, computing only the difference between consecutive frames to eliminate redundant bit-level operations, then amortizes data loading across multiple control steps via speculative inference, paired with a dedicated accelerator using a 1D systolic bit-serial PE array. It delivers up to a 34.2× speedup relative to a mobile GPU, with accuracy essentially unchanged.
Zheng Liu et al. · arXiv 2608.04428 source
PRIMAL3: Pushing Multi-Agent Pathfinding to City-Scale, 100,000 Agents · autonomy
Multi-agent pathfinding failures cluster at topological choke points — bottlenecks, dead ends, and locations of repeated conflict — where decisive coordination is essential. PRIMAL3 constructs features from cut vertices, dead-end regions, shortest-path distances, and blockage estimates, modeling same-direction following and opposite-direction conflict via separate graphs; during training, policy entropy identifies uncertain agents, LaCAM3 intervenes on trigger and supplies label-smoothed imitation targets, and LaCAM3 is no longer needed at inference. The approach scales to city-level deployments of 100,000 agents and was feasibility-tested on a real robot system.
Chengyang He et al. · arXiv 2608.04905 source
Talk2Sensors: First 3D Visual Grounding Dataset Including 4D mmWave Radar · perception
3D visual grounding (identifying a target in 3D space based on a language description) has largely been confined to indoor settings; outdoor extensions have mostly relied on monocular images alone, even though in real outdoor perception, cameras, LiDAR, and 4D radar each capture complementary texture, geometry, and motion properties. Talk2Sensors is the first multi-sensor 3DVG dataset built on camera + LiDAR + 4D radar, containing 8,682 language instructions and 20,558 referenced targets, with prompts deliberately aligned to physical cues specific to each sensor. The accompanying TSFormer dynamically routes appearance, geometry, and motion cues based on query semantics, outperforming the strongest baseline by 8.05 mAP on this dataset.
Runwei Guan et al. · arXiv 2608.04568 source
Other papers today: Faster-WAM, computing future representations once and sparsely reusing them throughout action denoising (arXiv 2608.04404 source); SAFECAST, improving training and calibration of VLA latent-state risk probes via contrastive-set perturbations (arXiv 2608.04246 source); GUARD, detecting failures by measuring diffusion VLAs' visual-language grounding degree via ablation, without modifying the pretrained policy (arXiv 2608.04510 source); CofactVLA, using counterfactual intervention to resolve VLA's "vision overriding language" confound (arXiv 2608.04396 source); Retrieve in Time, Correct in Frequency, a training-free test-time correction framework requiring no parameter updates or repeated inference (arXiv 2608.04527 source); RORA, end-to-end reconstruction of articulated simulation assets from a single static object video (arXiv 2608.04842 source); SpikingNav, a spiking neural policy for embodied navigation with lower parameter count and per-step compute, more robust under visual corruption (arXiv 2608.05078 source); EventKitchen, a large-scale stereo event-camera benchmark of kitchen human activities (arXiv 2608.04865 source); SSC, converting long-horizon manipulation demonstration sub-task annotations into automatically verifiable state-transition chains (arXiv 2608.04425 source).
Open Source · Tools · Benchmarks
· MAGI-2 Preview: Sand.ai open-sourced a MoE unified audio-video generation model with roughly 114B total parameters but only about 6B activated per forward pass, releasing weights, code, and training system together; it ranks 6th globally on the Artificial Analysis text-to-video leaderboard — video generation backbones are precisely the foundation underpinning today's world action models source
· Orbbec (Chinese 3D vision company) Robot-Free Data Collection Platform: Debuted at ROSCon JP 2026 in Japan, covering egocentric capture, hand-object interaction, and near-field wrist capture via its EGO, UMI, and WristCam lines; the demonstrated Dual-Ego prototype uses dual 2-megapixel fisheye cameras, sub-1ms sync precision, a six-axis IMU up to 1000 Hz, and weighs 200 grams total source
II. Funding and Deals
Unitree Robotics (Chinese humanoid maker) | STAR Market IPO Pricing | RMB 150.80/share | RMB 60.993 billion issue-date market cap · humanoid
Following the launch of preliminary book-building on August 5 (when market estimates put the price around RMB 104/share and market cap above RMB 40 billion), Unitree Robotics set its final IPO price at RMB 150.80/share, implying a listing market cap of about RMB 60.993 billion — notably above book-building estimates. The offering covers 40.446434 million shares, or 10% of post-issue total share capital, expected to raise roughly RMB 6.099 billion gross, or about RMB 5.917 billion net of issuance costs. The company proactively flagged valuation risk in its filing: the price implies a diluted P/E of 219.23× based on the lower of 2025 pre- or post-non-recurring-items earnings, and a static P/S of 35.89×, both above comparable industry levels. Total strategic placement shares are 8.089286 million; Hangzhou DeepSeek, controlled by Liang Wenfeng, received a placement allocation of over RMB 140 million with a 36-month lock-up, grouped alongside Tencent-affiliated Shanghai Qishan Investment under the category "strategic partnership or long-term cooperation intent with the issuer"; also in this tier are China National Petroleum Corporation's Kunlun Capital, China Southern Power Grid's industrial-financial holding arm, and China Telecom Capital. The company explained that the purpose of bringing in these investors is joint R&D focused on large AI models and embodied intelligence. Unitree also disclosed expected H1 2026 revenue of RMB 1.052–1.128 billion and net profit attributable to parent of RMB 258–306 million; subscription opens August 10.Sources: Beijing Daily App source; Sina Finance source
Hadrian (US) | New Round | Nearly $1.4 Billion | $7.9 Billion Valuation · industrial
Baillie Gifford and J.P. Morgan's Strategic Investments Group led the round, with Washington Harbour Partners and others participating. Hadrian builds highly automated factories for defense and aerospace customers; the valuation is more than quadruple its level in January this year, up from just $260 million in last year's round (which included Founders Fund). CEO Chris Power said the software platform driving its factories, Opus, is already in use by the US Army and Navy, and that the new funding will target capacity bottlenecks in the submarine, munitions, and drone industrial base. This round lands amid a dense stretch of defense-tech fundraising — Shield AI raised at a $12.7 billion valuation in March, autonomous-vessel maker Saronic at $9.25 billion, and Anduril's valuation more than doubled to over $60 billion two months later.Source: CNBC source
Avatar Robotics (US) | Seed Round | $6.5 Million · embodied
AlleyCorp led the round; the earlier pre-seed round was led by defy.vc, with Headline and others participating. The company builds warehouse humanoids around "remote human teleoperation plus progressive autonomy": robots perform picking, packing, kitting, sorting, and inventory tasks on customer sites while an operator remotely takes over from elsewhere, continuously generating data that feeds back into autonomous policy. Since launching in December 2025, its deployed robots have handled the packing and sorting of over 900,000 items, with customers including a top global beauty retailer, and it has entered a post-pilot expansion phase with a warehouse operator worth several billion dollars.Source: PR Newswire source
Strutt (Chinese robotics startup) | Pre-A+ Round | Nearly $100 Million Raised Cumulatively · autonomy
Founder Hong Xiaoping was formerly head of DJI's optoelectronics division and previously built the world's first mass-produced automotive-grade LiDAR. Its product, the Strutt EV1, takes the form of a wheelchair but is built around an L2-level autonomous driving core: two LiDARs, ten ToF sensors, six ultrasonic sensors, and two cameras, on an automotive-grade four-wheel-drive chassis with independent suspension on all four wheels and ABS, top speed 13 km/h and 32 km range. Overseas pricing is $7,499, with a planned Chinese price of RMB 29,999; overseas mass production and delivery have begun, with entry into the Chinese market planned for H2 2026. The company says the return rate on its first delivery batch was zero, and the test-drive-to-paid-purchase conversion rate is nearly 50%.⚠️ Single-party accountSource: Robotics Lecture Hall source
Shelfmark (Pittsburgh, US) | Seed Round | $3.5 Million | ~$5 Million Raised Cumulatively · industrial
Armory Square Ventures led the round, with Grand Ventures, Hyde Park Angels, and others participating. The company provides in-line quality inspection for continuous-flow manufacturing — production lines making industrial film, webbing, paper, flooring, and coiled metal at high speed and volume, where defects often aren't caught until large batches of material are already wasted. Its platform packages in-line industrial cameras, spatial sensing, and a proprietary deep-learning vision model into a managed service, and causally correlates defects with environmental variables like temperature, humidity, and pressure; the company says its customer deployments achieve 99.5% defect-detection accuracy and roughly a 50% reduction in manual inspection labor costs.⚠️ Vendor accountSource: Pulse 2.0 source
III. Commercialization and Deployment
Japan's First Hospital-Validated Commercial Humanoid Goes on Sale, Built on a Unitree G1 Chassis · humanoid
Tokyo AI company ZEALS's subsidiary Omakase Robotics launched D1 for sale on August 5 at ¥5 million (about $32,000) per unit — Japan's first humanoid robot to pass a hospital proof-of-concept and go on public sale. The design reflects compliance-driven tradeoffs throughout: standing 129.3 cm tall with a base roughly 48 cm wide, sized to pass through standard door frames, fit into elevators, and navigate narrow hospital corridors without building modifications; it uses a wheeled base rather than bipedal legs — per Omakase CTO Shuichi Nagao's remarks at the Tokyo Humanoids Summit this past May, bipedal humanoids currently cannot be sold in Japan for safety and regulatory reasons; each arm joint has a torque sensor that stops the robot instantly on contact with a person or object; battery life is about 8 hours, matching a standard work shift. Notably, it runs on a Unitree G1 chassis, a platform that was added to the US FCC's Covered List on security grounds just 8 days before D1's commercial launch — what a "quasi-domestic" product means when its supply chain still runs through a Chinese manufacturer is a question now facing the Japanese hospitals being asked to sign deployment agreements.Source: Tech Times source
Video Action Models Enter an Audi Plant: Robots Begin Learning Directly from Watching Humans · embodied
The video action model FLUX-mimic, released jointly by German company mimic robotics and Black Forest Labs, has entered an Audi factory environment for application validation. The approach is to understand human operating motions from video and translate them into executable robot behaviors, targeting assembly, loading/unloading, flexible grasping, and quality inspection tasks that aren't fully structured — tasks that traditionally require engineers to repeatedly demonstrate and hand-write rules, and to redo the tuning whenever the model or station changes. Hard constraints for industrial deployment remain unsolved: factories need deterministic, traceable actions with safety margins, and the model can't just perform well on average — force feedback, simulation validation, and safety controllers are all still required.⚠️ Application validation stageSource: AGV.com.cn (reprinted via Sina Finance) source
Zoox's Steering-Wheel-Free Robotaxi Begins Paid Rides in Las Vegas August 10 · autonomy
Following last week's approval of the first US commercial exemption for a steering-wheel-free robotaxi (previously reported), Zoox has set August 10 as the date it begins charging for rides, starting in Las Vegas. Since opening free rides across its test cities in September 2025, it has accumulated over 500,000 rides, with a waitlist of comparable size. Miami and Austin are its next named destinations.Source: TechCrunch
PIA Automation Builds E-Drive Assembly Line for BMW's Steyr Plant · industrial
The project is located at BMW Group's Steyr plant in Austria, spanning two floors totaling over 5,600 square meters, with 78 processing stations, 27 robotic cells, and 46 industrial robots — one of PIA's largest projects to date. The line first assembles the EV transmission, then "marries" it with the motor in a robotic cell where 12 synchronized electric screwdrivers tighten it in one pass into a complete E-Drive unit; automatic gripper changeover lets different transmission variants switch without manual intervention, supporting single-piece batches. The entire process is validated in advance using digital twins and virtual commissioning of axes, sensors, material flow, and on-site equipment behavior; the final sections of the line are scheduled to enter commissioning in 2027.Source: Robotics & Automation News source
Nucleus Emerges from Stealth: Humanoid Robot Into a German Factory in 90 Days · humanoid
European robotics startup Nucleus came out of stealth on August 5, saying it deployed a humanoid robot to the site of a leading German industrial manufacturer in under 90 days; the robot performs factory tasks under human supervision while accumulating data to feed subsequent autonomy development. The founding team comes from CERN, the European Space Agency, 1X, NEURA Robotics, and Agile Robots. The customer's name, robot unit count, and task scope were not disclosed.⚠️ Single-party accountSource: Interesting Engineering source
IV. Industry Developments
GigaDevice (Chinese chipmaker) Launches Two Robot-Specific MCUs, Bringing EtherCAT On-Chip · hardware
GigaDevice released the GD32H77R robot-specific chip and the GD32F50MxxG highly integrated motor control chip, targeting humanoid joint modules/servo drives and lightweight servo joints respectively. The GD32H77R is a 600 MHz Arm Cortex-M7 with 640 KB of tightly coupled memory running at CPU frequency for zero-wait-state instruction fetch; it integrates an EtherCAT slave controller and dual 100 Mbps Ethernet PHYs directly on-chip, eliminating two external industrial-grade PHY chips, achieving one-way transmission latency as low as under 50 ns and DC synchronization cycle precision of 62.5 μs; the company says dynamic power consumption is only 20%–50% that of comparable-architecture M7 competitors. The GD32F50MxxG integrates a proprietary three-phase gate driver and a four-channel op-amp into the same package, eliminating the need for external standalone drivers and sampling op-amps to simplify BOM; the 8mm×8mm QFN80 variant enters mass production this December, and the 7mm×7mm BGA100 variant ships in volume next March. A company marketing representative said humanoid-robot-related MCU shipments reached roughly 3 million units in H1 this year, with volumes expected to grow further in H2. Shares closed up 8.26% at RMB 385.44 that day, with a market cap of RMB 270.481 billion.⚠️ Vendor accountSource: ChipInsight source
South Korean Government Sets Physical AI Direction: Deploying 1,000 AI Robots Per Year · industrial
South Korean Deputy Prime Minister and Minister of Economy and Finance Koo Yun-cheol chaired an emergency economic headquarters meeting and an economic structural reform ministerial meeting on August 6, announcing the "K-Asset Innovation Plan" and economic restructuring direction. The robotics-related portion includes: advancing AI process innovation for core industries like steel, petrochemicals, and shipbuilding, developing purpose-built humanoid robots for 10 industries, and pushing an annual deployment scale of 1,000 AI robots, aiming to gain first-mover advantage in physical AI. The same package includes a comprehensive overhaul of the 1950 State Property Act — historically focused on real estate — into a "Basic National Asset Act" that expands the scope of the already-1,402.7-trillion-won national asset base to include securities equity, intellectual property such as patents, and virtual assets.Source: Seoul Economic Daily source
Pony.ai Gives Robotruck Mass-Production Timeline · autonomy
Pony.ai targets deploying 500–1,000 unmanned heavy trucks over the next 2–3 years, with unmanned light trucks aiming for a fleet of 100,000 vehicles before 2030. Verifiable facts: the trucking business contributed nearly half of the company's revenue in 2025; the company obtained China's first nationwide permit in April this year for driverless-following operation in autonomous truck platooning. The company says its robotaxi, unmanned light-truck, and heavy-truck businesses share over 95% synergy in driverless hardware, policy frameworks, and supply chains, and that unmanned pure-electric light trucks can cut per-kilometer freight costs by 40%–50%. The industry backdrop is a driver shortage — the "2025 China Road Freight Industry Talent Survey Report" says the number of truck drivers nationwide fell by roughly 3.8 million from 2019 to July 2025, with those under 40 accounting for less than 10%.⚠️ Plan-stage figuresSource: Caijing Tianxia WEEKLY (reprinted via Sina Finance) source
LG Chairman to Meet Jensen Huang Again in Silicon Valley Next Week, Agenda Points to Robotics Cooperation · adjacent
According to multiple South Korean outlets, LG Group Chairman Koo Kwang-mo will travel to Silicon Valley next week to meet with Nvidia CEO Jensen Huang, discussing concrete cooperation plans in AI infrastructure and robotics — about two months after the two met in Seoul this past June. Around the same time, Nvidia is also deepening AI and robotics cooperation with Hyundai Motor Group. Details and outcomes of the meeting have not been announced.Source: Yonhap News Agency source
DOBOT Unveils Consumer Humanoid DOBOT LUMO · humanoid
DOBOT unveiled a humanoid robot, LUMO, marketed around "embodied all-terrain" capability, standing nearly 1.3 meters tall; its promotional video shows continuous walking across grass, sand, and cobblestone terrain, plus outdoor sparring, home companionship, and dance-mode switching. The company positions it as a consumer-grade companion robot and has not disclosed a mass-production timeline or pricing. The showcased content is a demo video, not a mass-production or scaled-delivery metric.⚠️ Vendor accountSource: 36Kr source
Hardware · Supply Chain
· PaXini PX-FOOTRIX: Released a multi-dimensional plantar tactile sensor based on 6D Hall-array tactile sensing technology, providing full-area 3D array force sensing and six-axis force/torque sensing on the sole of the foot, supporting customizable IP67/68 protection with claimed 1000% shock overload tolerance, and compatible with SPI/UART/RS485/CAN FD interfaces; the company calls it "the world's first" and says it will debut at WRC 2026 ⚠️ Vendor account source
· Molex Hybrid Connector: Launched a compact hybrid connector for humanoid robots and industrial automation, addressing the longstanding problem of limited wiring space inside joints source
· GAM Enterprises × Schaeffler: GAM will manufacture PSC planetary reducers for Schaeffler — after Schaeffler recently placed orders on the finished-humanoid side in Europe, upstream capacity is now beginning to look outward for contract manufacturing source
· Horizon Robotics (Chinese embodied-AI chipmaker) × GigaDevice: The two companies jointly released a Chinese-made "large brain/small brain" solution for six-axis robotic arms, integrating the controller and main control chip into one domestic supply chain source
V. This Week's Watch
California Truckers' Union Sues DMV, Seeks to Overturn New Driverless Heavy Truck Rule · autonomy
Teamsters California filed suit against the California DMV in Alameda County Superior Court on August 5. The dispute centers on procedure: the DMV first adopted the rule in April this year, allowing commercial vehicles over 10,000 pounds, including trucks, to operate fully autonomously on California public roads. The union alleges the DMV used an expedited rulemaking process intended only for regulations whose first-year implementation costs or benefits fall below $50 million, in order to bypass legally required economic-impact studies and public disclosure — and that in its filing the DMV claimed the rule would not eliminate any jobs or businesses, even though reducing demand for commercial truck drivers is itself an openly stated goal of autonomous freight companies. The complaint also alleges the DMV failed to adequately consider the safety risks of sharing roads with driverless trucks that are still being tested and have not been fully validated, and asks the court to find the rulemaking process fundamentally flawed and to vacate the rule. This marks the first time in the US that a driverless heavy-truck approval process has been challenged in court at the state level, and the outcome will directly affect the timeline for California, the largest testing ground for the technology.Source: Courthouse News Service source
LexisNexis's First Humanoid Robot Patent Report: China Leads on Volume, the US on Per-Patent Quality · humanoid
Legal and IP information firm LexisNexis released the first global innovation report on humanoid robots, covering over 26,000 patent families across three technology areas: interaction systems, body structure design, and control/planning architecture. On volume: as of 2025, China holds 73% of body-structure-design-related patents, versus South Korea 11%, Japan 7%, and the US 5%. Switching to a weighted patent asset index accounting for technical relevance and market coverage, China still leads at 63%, followed by the US at 11%, Japan at 10%, and South Korea at 7% — but the US ranks first in average per-patent competitiveness, a conclusion only visible when volume and quality are examined separately. Over the past decade, China's patent activity has grown roughly 5x in control/planning architecture and about 2.5x in body structure and 2x in interaction systems. In terms of patent-holder structure, companies account for 92%, 83%, and 86% of patents in Japan, Germany, and the US respectively, versus 58% companies and 39% research institutions in China. The top ten by patent asset index are Fanuc, Alphabet, UBTECH (Chinese humanoid robotics maker), Kawasaki Heavy Industries, the Chinese Academy of Sciences, Nvidia, Strong Force Innovation, Harbin Institute of Technology, Intuitive Surgical, and Samsung, with nine of the ten having humanoid-related patents accounting for under 10% of their total patent portfolio — UBTECH is the only specialized company exceeding 25%. The startup ranking's methodology deserves note: it evaluates only unlisted companies that have publicly demonstrated a bipedal-walking humanoid, excluding Tesla, Boston Dynamics, and the already-listed UBTECH — under this criterion the top five are all Chinese, in order: Fourier Intelligence (Chinese humanoid maker), Agibot (Chinese humanoid startup), Limx Dynamics (Chinese humanoid startup), Pudu Robotics (Chinese service robot maker), and Unitree, with AgiBot (LeJu Robotics) ranking 9th. The report's cited market forecasts diverge sharply: Grand View Research and Goldman Sachs project roughly $40 billion by the early-to-mid 2030s, while Morgan Stanley projects the figure, including supply chain and maintenance, could exceed $5 trillion by 2050.Source: Robotics Outlook (reprinted via ifeng Tech) source
Production Racing Toward 100,000 Units, But Under 5% Actually Reach the Factory Floor · humanoid
Gan Xiaobin, Deputy Director of the Ministry of Industry and Information Technology's Science and Technology Department, gave figures at WAIC in July this year: China's humanoid robot production was roughly 20,000 units in 2025, already over 40,000 units in H1 2026, and is expected to surpass 100,000 units for the full year. A more noteworthy set of figures from the same period comes from industry statistics: China's market sold roughly 23,000 humanoid robots in H1 2026, with about 30% going to research institutions, 30% used for trade shows, over 30% exported, and under 5% actually entering factory production lines. Taken together, these two figures show that production growth and depth of industrialization are not currently the same thing — the boom in shipment figures has not yet translated into shop-floor hours.⚠️ Third-party compilationSource: AI Research Notes Pro source (WeChat, CN)
This Week's Supply Chain Roundup: The Cost-Reduction Curve Is Bottlenecked at Transmission Components, With Tightest Supply in Harmonic Reducers and Lead Screws · hardware
This week's in-depth component coverage converges on the same conclusion: the main battlefield for cutting humanoid costs isn't in algorithms or the finished robot, but in transmission components. On a comparable basis, Optimus's bill-of-materials cost is about $131,000 without relying on the Chinese supply chain, dropping to about $46,000 when it does — nearly a threefold gap; actuators, lead screws, and reducers together account for about 68% of its BOM. By component: Chinese-made harmonic reducer prices run 40%–60% of Harmonic Drive's (Japanese maker), with Chinese maker Leaderdrive's unit price falling from about RMB 1,900 in 2017 to about RMB 1,000 in 2025, with potential to reach RMB 400–500 at scale. The planetary roller screw has the highest barriers — each unit needs 10–14 screws, priced RMB 5,000–10,000 imported versus about RMB 2,000 once localized — with the bottleneck being 0.5-micron-level machining precision and raising yield from 60% to 85%; foreign suppliers still hold over 70% of this market. Six-axis force sensors are about 55% localized, with each unit requiring 4 six-axis sensors plus 28 joint torque sensors.
The tight spots are equally specific. Per Soochow Securities, global humanoid harmonic reducer demand is projected at roughly 8.8 million units by 2027, against effective capacity of only about 5 million units — a shortfall exceeding 20%. Leaderdrive signed a three-year exclusive agreement with Unitree covering reducers for 110,000 finished robots, with orders booked through Q3 2027 and capacity utilization around 95% — yet its gross margin has already fallen from 41% in 2023 to 33.6% in Q1 2026, as capacity expansion and price competition happen simultaneously. Lifespan is another unresolved hurdle: planetary roller screws require precision held within ±6 microns while joint service life is only about 1 year, and precision components like reducers generally last around 8,000 hours — under an 8-hour workday, that's under three years before replacement is needed, still short of the 5+ years industrial applications require. Tesla's guidance to suppliers targets 1,000 units/week by September and 2,000–2,500 units/week by year-end (equivalent to a year's supply of components for 100,000 units), though several supply-chain sources believe the year-end target won't be met; Elon Musk himself said on the Q2 earnings call on July 22 that Optimus will be the hardest product Tesla has ever mass-produced, citing the fact that "there is no supply chain."⚠️ Brokerage research compilationSources: Tech Dissection source (WeChat, CN); AI Research Notes Pro source (WeChat, CN)
Top comments (0)