Today's Highlights
· SoftBank exercises its put option, Hyundai buys out Boston Dynamics' last 9.65% stake for $325 million
· Nvidia teams up with the Japanese government to build the world's first national-level physical AI compute base, with 27,500 Rubin GPUs
· RSS 2026 Best Paper FlashSAC: humanoid locomotion sim2real training compressed from hours to minutes
· Harmonic reducer prices halve in three years, bulk quotes down to just over 2,000 yuan
· NHTSA to finalize new driverless-vehicle rules by 2028, replacing hardware-design clauses with a "behavioral capability" test
I. Paper Progress
FlashSAC: compressing humanoid locomotion sim2real training from hours to minutes · locomotion
RSS 2026 Best Paper. The conference this year accepted 210 of 708 submissions, a 29.7% acceptance rate, with 8 papers making the finalist list. The work tackles an old problem in off-policy reinforcement learning for high-dimensional robot control — slow, unstable training. It goes against convention by reducing the number of gradient updates while simultaneously scaling up model size, data throughput, and the experience replay buffer, then controlling value-network error accumulation by bounding weight, feature, and gradient norms. It outperforms PPO and various off-policy baselines overall across 10 simulators and 60-plus tasks, and the most visible result is a humanoid locomotion sim2real experiment where training time shrank from hours to minutes. The authors are from Holiday Robotics, KAIST, KRAFTON, TU Darmstadt, and KTH — Holiday Robotics is the same company that set a new Korean Series A record of KRW 155 billion just last week.
Donghu Kim, Youngdo Lee et al. (Holiday Robotics × KAIST × KRAFTON) · RSS 2026 Best Paper · Analysis: QbitAI source (WeChat, CN)
NeuralActuator: giving low-cost robot arms a sense of how much force they're feeling, without torque sensors · perception
RSS 2026 Best Systems Paper. Friction, backlash, and temperature drift throw off the traditional current-to-torque relationship — a major source of sim2real gap for cheap robot arms. This work uses a transformer combined with differentiable simulation to jointly predict actuator torque, external contact force, contact probability, and motor state, letting robots sense applied force at deployment without dedicated force/torque sensors. It was validated on three classes of robot arms ranging from about $500 to over $30,000, and the team also released the NAD dataset collected via bimanual teleoperation.
Zhiyang Dou et al. (MIT CSAIL) · RSS 2026 Best Systems Paper · Analysis: QbitAI source (WeChat, CN)
DAPL: robots learn to "use the environment as leverage," extracting objects without grasping · manipulation
Selected for the RSS 2026 finalist list. The framework first uses a world model to learn the dynamics of collisions, sliding, and tipping among objects in cluttered scenes, then uses these dynamics representations to guide reinforcement learning — so the robot no longer relies solely on direct grasping, and can also complete "extrinsic dexterous manipulation" like pushing, nudging, and flipping by making contact with surrounding objects and the environment. In unseen simulated cluttered scenes, success rate improved by more than 25 percentage points over baselines such as grasping and teleoperation, and reached about 50% across 10 categories of real-world scenes; it has already been used for shelf-picking tasks. The authors emphasize these behaviors emerged autonomously from the model rather than being tuned via reward hacking.
Yixin Zheng, Jiangran Lyu et al. (CASIA × BAAI × Galbot × Peking University × SJTU) · RSS 2026 Final List · Analysis: QbitAI source (WeChat, CN)
UniTac: letting VLA "imagine" the feel of touch before contact · manipulation
The difficulty in bringing touch into VLA isn't a lack of data, but that the data is locked inside heterogeneous sensors — the same object can produce completely different signals across different tactile sensors. UniTac splits touch into two separately modeled parts — "object physical properties" and "sensor configuration" — and learns a unified representation from over 400,000 publicly available tactile videos, then predicts latent tactile states from vision and feeds them into VLA, shifting touch from post-contact feedback to a pre-contact prior. On the understanding side it scores 66.51 on PHYSICLEAR-Test, ahead of models like Octopi; on the generation side it achieves an average SSIM of 0.836 across four sensor types. The value lies in scenarios like paper cups or towels — objects that "break if you squeeze wrong" — where correcting after contact is already too late.
UniX AI × Zhejiang University × MIT × Oxford × Yale × SJTU · Accepted at ECCV 2026 · Analysis: QbitAI source (WeChat, CN)
GigaWorld-Policy-0.5: a world-action model that no longer generates video at inference time · world-model
21↑ on Hugging Face trending. Existing world-action models (WAMs) generally generate explicit future video at inference time, and the compute overhead is heavy enough to break real-time closed-loop control. This version takes an "action-centric" approach — using dense supervision from future visual dynamics during training, decoding only actions at inference, and introducing a Mixture-of-Transformers to split visual dynamics modeling and action generation across different experts, achieving 85ms inference latency on a single RTX 4090. The team also used an agentic AutoResearch pipeline to automatically search training configurations, cutting out a large amount of manual tuning.
GigaWorld Team, Angen Ye et al. · arXiv 2607.13960 source
WANDA: synthesizing a trainable world from a single real-robot demonstration · benchmark
The labor cost of teleoperation and UMI-based data collection simply doesn't scale for open-world mobile manipulation. WANDA's approach is to squeeze maximum value from a single demonstration: it first reconstructs a background Gaussian splat and robot-object interaction trajectories from RGBD observations as a "world substrate," then rearranges contact-rich interaction segments across many spatial configurations and strings them into new trajectories via whole-body motion planning, synthesizing realistic observations across diverse 3D worlds generated from everyday photos. Experiments show policies trained this way simultaneously gain long-horizon robustness, spatial generalization, and cross-environment generalization, and naturally support cross-embodiment data generation — it has already been zero-shot deployed on another mobile manipulation robot with a different form factor.
Lingxiao Guo, Huanyu Li, Guanya Shi · arXiv 2607.13154 source
APT-RL: quadrupeds autonomously switch gaits for high-speed traversal in the wild, using only onboard sensors · locomotion
The difficulty in multi-skill locomotion lies in the handoff between skills — when to switch gaits and to which one is usually decided by a hand-crafted state machine. This framework uses an action-pretrained transformer to unify multi-skill locomotion, enabling autonomous skill switching and high-speed traversal across complex terrain using only onboard perception and compute, validated from rugged wilderness to urban environments.
Jun-Gill Kang, Jaehyun Park et al. · arXiv 2607.13579 source
Other papers today: M4World, a multi-view multimodal driving world model supporting object-level interaction and minute-scale streaming generation (arXiv 2607.14005 source); Kepler-Encoder-v0.1, treating proprioceptive state as a modality fused with vision and force/torque into a shared latent space — vision-only features reading force have an R² as low as 0.10 (arXiv 2607.13522 source); Anchor-Align, using a frozen VLM copy for layer-wise distillation to prevent behavior-cloning fine-tuning from overwriting pretrained representations (arXiv 2607.13429 source); Semantic Anchoring, anchoring action representations back to a semantic manifold and splitting them into shared-semantic and private channels (arXiv 2607.13597 source); Reverse to Advance, exploiting a task's directional asymmetry, using an "easy reverse task" to lower teleoperation collection cost for hard tasks (arXiv 2607.13455 source); PhysClaw-0, letting language corrections persist and be reused across turns so supervision cost scales with the number of problem types rather than conversation length (arXiv 2607.14047 source); UniPhysGen, automatically converting raw 3D assets into simulation-ready assets with articulation semantics and physical properties (arXiv 2607.13586 source); From Pixels to States, examining interactive world models through four game-engine dimensions, with a data engine featuring over 90 hours of "Black Myth: Wukong" data (arXiv 2607.14076 source); Grounded world models, proposing five reference points for embodied world models drawn from biological neural circuits (arXiv 2607.13560 source); showing the SIGReg objective is equivalent to variational free energy, providing JEPA world models with a normative active-inference interpretation (arXiv 2607.13612 source).
Open Source · Tools · Benchmarks
· X-TouchMind V1 and TacVerse 1k: Qianjue Robotics released what it says is the industry's first tactile-focused VTLA embodied foundation model, paired with a 1,000-hour native visual-tactile dataset — 100% tactile data coverage, with vision, touch, force, pose, and high-frequency dynamics captured together rather than tactile data added on after the fact to vision-only data; the model uses a System 0-2 hierarchy, with System 0 dedicated to high-frequency tactile correction source (WeChat, CN)
· RLinf v0.3: an update to the embodied reinforcement learning framework jointly built by Infinigence AI and Tsinghua University, with five capability upgrades spanning model ecosystem to real-robot deployment source
· GPUSimBench: a benchmark for GPU-accelerated simulators establishing standards for scalability, physical consistency, and computational determinism, pointing out that unchecked scaling sacrifices reproducibility (arXiv 2607.13059 source)
· Industrial Dexterity Benchmark: an integrated hardware-software benchmark platform for industrial dexterous manipulation, covering cable routing, connector insertion, and precision assembly — tasks still done by hand today (arXiv 2607.14021 source)
· HRIBench: writes collaborative tasks as executable scripts with explicitly modeled roles, temporal dependencies, and coordination constraints, filling a near-blank spot in VLA evaluation for human-robot interaction structure (arXiv 2607.13056 source)
II. Funding and Deals
Boston Dynamics | Equity Buyout | ~$325 Million | 9.65% Stake · humanoid
Following yesterday's reports that Hyundai planned to buy out SoftBank's holding, the trigger and price of the deal were confirmed today: SoftBank actively exercised its put option, selling its remaining 9.65% stake back to Hyundai Motor Group for about $325 million, giving Hyundai 100% ownership. The price is worth a closer look — when Hyundai bought 80% in 2021 for 1 trillion won (about $674 million), Boston Dynamics was valued at about 1.24 trillion won; that valuation has now surpassed 30 trillion won. At the current valuation, 9.65% should be worth roughly $2 billion, meaning SoftBank exited at a discount under the option terms agreed years ago. Hyundai also laid out a timeline for Atlas: entering its Metaplant America factory in Georgia for parts sequencing in 2028, expanding to component assembly by 2030. Multiple Korean media outlets also noted that consolidating ownership clears the way for a Boston Dynamics IPO in the US.Source: Yonhap source | Bloomberg source
Senior Auto (Chinese autonomous trucking startup) | Series C | RMB 300 Million · autonomy
Jointly invested by Xingzheng Capital and Yidao Capital, the funds are primarily for R&D on next-generation production-grade autonomous driving solutions. Founded in 2020, this driverless heavy-truck company follows a "closed → semi-open → open" progressive rollout, monetizing through operating-as-a-service and technology licensing, and has already achieved scaled commercialization in ports and closed parks. Its product is centered on world-model-driven decision-making, covering onboard intelligence, fleet management, and cloud dispatch. Around the Yiyongzhou corridor, it has connected a complete chain across three node types — Ningbo-Zhoushan Port, the Yiwu Suxi rail yard, and the Yongjin Expressway — spanning port drayage, yard transfer, and trunk-line transport. This approach of "using cash flow from closed scenarios to sustain the push into open scenarios" is one of the few autonomous heavy-truck business models in China currently proven to work.Source: PEdaily source
microagi (Germany) | Seed Round | $55 Million · embodied
Led by Hummingbird, with participation from Northzone, LocalGlobe, Village Global, and redalpine — the largest seed round in German startup history. Founder Bercan Kilic left a job as an aerodynamics engineer at Red Bull Racing shortly after landing it in 2023, and started the company last year out of a hacker house in Munich. The company neither builds robots nor foundation models; instead it records workers doing their jobs using cameras and sensor gloves, then uses that data to fine-tune off-the-shelf robot models into specialists for a customer's production line. Five companies currently contribute data on its platform, with one preparing for deployment. Its consumer-facing offshoot, shift, went viral earlier this year by offering "free housework in New York in exchange for first-person recordings," and this week started offering free private chef services in San Francisco.Source: Business Insider source
Think Surgical (US) | Debt Financing | Up to $65 Million · adjacent
The orthopedic surgical robotics company reached an agreement with healthcare investment firm Symbiotic Capital worth roughly RMB 440 million, to advance commercialization of the handheld micro-orthopedic robot TMINI. Choosing debt over equity typically signals the company places a higher priority on avoiding dilution and has confidence in predictable existing cash flow.Source: MedRobot source (WeChat, CN)
SwitchOn (India) | New Round | $8 Million · industrial
A physical AI company focused on manufacturing visual quality inspection, funding will go toward scaling production-line inspection deployments.Source: Inc42 source
III. Commercialization and Deployment
Machina Labs wins Lockheed Martin JASSM missile qualification contract · industrial
This marks the first time components produced by a Machina factory have advanced to qualification testing for a defense missile system — a higher bar than a general "partnership intent." The company's proprietary RoboCraftsman cells autonomously handle loading, metal forming, scanning, trimming, and drilling, turning digital designs directly into production-grade metal parts; the delivered component is a fuel tank metal structure, produced at its Los Angeles factory using its RoboForming process combined with precision laser welding, and handed to Lockheed Martin for testing and system integration. Follow-on production volume will be handled by the under-construction Factory 3: 200,000 square feet, capable of housing up to 50 RoboCraftsman cells. The key bottleneck for robots serving critical structural parts in defense supply chains is replacing tooling and dies with seven-axis flexible forming — once this qualification hurdle is cleared, the marginal cost of switching to a new part becomes a different order of magnitude from traditional stamping.Source: The Robot Report source
Tesla Robotaxi registered vehicles in Texas rise to 175, up 58 in a single day · autonomy
According to a Texas autonomous-vehicle tracking website, Tesla's registered autonomous vehicles in the state have reached 175, with 58 added in a single day — a marked acceleration in pace. An important caveat: this is the registered count, not the actual number in active operation — Electrek reporters previously conducted an on-the-ground count across thousands of square miles of the Austin metro area and found the number of vehicles actually operating far below the registered total. On the other hand, one data point favors Tesla: according to NHTSA incident reports, its Austin fleet has not reported a single at-fault accident since mid-April this year; the only new incident was a Model Y that was rear-ended while parked.Source: Autohome source | LeiPhone source (WeChat, CN)
Stryker's Mako orthopedic surgical robot power system goes fully commercial in the US, cumulative surgeries surpass 2.5 million · adjacent
The system, aimed at knee replacement, has moved from a limited release to full US commercial launch. The same day, the company disclosed that its Mako platform has completed over 2.5 million surgeries across 47 countries — a solid data point that surgical robotics has already cleared the commercialization-validation bar, and also the real benchmark that general-purpose humanoid research results — like papers on Unitree's G1 performing surgery — should actually be measured against.Source: MassDevice source | Stock Titan source
DHL Malaysia uses AI drones for warehouse inventory checks, cutting cycle time by 92% · industrial
Drones automatically perform cycle counts inside the warehouse, cutting inventory-check time by 92%. Warehouse counting is a textbook case of "high-frequency, low-skill, extremely hard to staff" work, and one of the easiest categories for automation ROI to pencil out clearly.Source: BusinessToday Malaysia source
Ukraine's Nurse TB20 ground robot enters mass production · adjacent
This ground unmanned platform has begun mass production, mainly for battlefield casualty evacuation and resupply. Wartime demand is compressing the "prototype to mass production" cycle for ground robots well outside the normal defense-industry pace.Source: UA.NEWS source
IV. Industry Developments
Nvidia and the Japanese government build the world's first national-level physical AI compute base · hardware
Nvidia announced a partnership with Noetra to build a Vera Rubin AI factory, configured with 13,750 Vera CPUs and 27,500 Rubin GPUs, 140 megawatts of data center capacity, using Vera Rubin NVL72 racks and the DSX platform, interconnected via Spectrum-X Ethernet. This compute will support Japan's METI-launched FRONTia project (developing multimodal foundation models for AI robotics and physical AI), feeding Japanese manufacturing know-how and real industrial data into the models; the pretrained weights Noetra trains will be opened up to Japan-based model developers and companies, alongside open models such as Nemotron, Cosmos, and Isaac GR00T. Announced the same day on the industry side: Japanese manufacturing giants Fanuc, Yaskawa Electric, Kawasaki Heavy Industries, and Toyota's Woven City said they will build their respective physical AI capabilities on NVIDIA Cosmos. Jensen Huang put it bluntly: "Japan invented modern manufacturing, and now it is building the AI factories that will power the next industrial revolution." Set alongside this week's Hyundai buyout of Boston Dynamics and Samsung's plan to build a humanoid base, the three East Asian manufacturing powers are each betting on physical AI according to their own asset structures: Japan is betting on infrastructure and open models, Korea on complete units and bringing embodiment in-house.Source: NVIDIA Newsroom source | Chosun Ilbo source
Nvidia launches Jetson Thor T3000/T2000, pushing robot compute into mainstream price points · hardware
The T3000 delivers 865 FP4 TFLOPS of compute at roughly half the size and power draw of the T5000, with a Blackwell GPU, an eight-core Neoverse Arm CPU, 32GB of LPDDR5X, and 273GB/s of bandwidth; despite the halved footprint, inference performance on multimodal workloads such as large language models, VLMs, VLA, and world foundation models remains comparable to the T5000. The entry-level T2000 offers 400 FP4 TFLOPS and 16GB. Nvidia pointed to a very concrete motivation right now: moving to the T3000 cuts costs amid rising memory prices — and newly released Jetson agent skills can automatically optimize memory usage, letting developers drop a memory SKU tier within the same product class without losing performance. The customer list already includes 1X, Agile Robots, Amazon Robotics, Boston Dynamics, Fanuc, Hitachi, Techman Robot, and UBTech. This extends Nvidia's edge platform lineup from 70 TOPS all the way up to 2,000 TFLOPS.Source: NVIDIA Blog source | VideoCardz source
WAIC unveils its "top ten exhibits," only one humanoid robot makes the list · humanoid
The 2026 World Artificial Intelligence Conference opens tomorrow, and its top-ten-exhibits list includes only one humanoid robot — Zhiyuan Robotics' (Chinese humanoid startup) full-size commercial-scenario model Yuanzheng A3 Ultra. Another related selection is Ant Group's Lingbo smart pharmacy, which emphasizes deploying robots without any store renovation under a "one brain, many machines" approach. The shift in judging criteria is more notable than the list itself: doing backflips no longer counts — what counts is running stably on production lines and in stores. Also from the show floor: the number of embodied-intelligence exhibitors rose from about 80 last year to over 200 this year, and roughly 60 full-size humanoid robots will be deployed at the venue for guiding and answering questions.Source: Cailian Press source | Purple Cow News source
Samsung Electronics reportedly plans KRW 13 trillion humanoid robot base in Gumi · humanoid
According to Korean outlet The Elec, the base would be overseen by Samsung Electronics' DX division, with the humanoid production line tentatively sited at an unused factory in Gumi's No. 1 Industrial Complex, while Samsung SDS builds a supporting AI data center locally. ⚠️ Reported by Korean media, unconfirmed Samsung has not publicly confirmed this, and both scale and timeline remain to be finalized. Samsung's existing positioning offers context: it has raised its stake in Korean robotics company Rainbow Robotics to 35% through investment and exercising call options, participated in a RMB 1 billion strategic round for Xingdong Jiyuan (Chinese humanoid startup) in March, and this week Samsung Ventures joined Walden Robotics' $300 million seed round.Source: WeChat Humanoid Robotics source (WeChat, CN)
Public consultation launched on industry standard for humanoid robot dexterous hands · embodied
The standard sets out a full set of technical requirements and standardized test methods for dexterous hands, covering degrees of freedom in appearance, motion performance, perception capability, grasping and manipulation capability, common device interfaces, electromagnetic compatibility, environmental adaptability, and operational reliability, along with accompanying test procedures — filling a regulatory gap in this segment. Dexterous hands are currently the highest-priced single component and also the one with the least cross-comparable specifications — every manufacturer quotes its own degrees of freedom, payload, and tactile channels differently. Once the standard takes effect, buyers will at least have a common yardstick.Source: Household Service Robot Committee source (WeChat, CN)
China adds 27 vocational-education majors including "humanoid robot engineering technology" · humanoid
China's Ministry of Education added 27 new majors to its vocational education catalog, with humanoid robot engineering technology among them. What actually bottlenecks factory-floor deployment at the ten-thousand-unit scale is often not the model, but people who can commission, maintain, and script the workflows — a change to the major catalog signals the supply side is beginning to gear up on an industrialization timeline rather than a research one.Source: China News Service source
Nvidia reshuffles autonomous-driving leadership, Wu Xinzhou (Nvidia autonomous driving VP) consolidates authority · autonomy
According to an exclusive from 36Kr, Nvidia carried out a major personnel reshuffle in its autonomous driving business, with authority consolidating around Wu Xinzhou. This follows yesterday's reported hire of a former Qualcomm autonomous-driving VP — the convergence of "ex-Qualcomm" and "ex-Xpeng" factions within Nvidia's automotive unit is becoming clearer.Source: 36Kr source (WeChat, CN)
Hardware · Supply Chain
· D-Robotics Sunrise S600: this high-compute chip aimed at embodied intelligence has entered mass-production validation, endorsed by over 20 leading embodied-intelligence companies (carried over from yesterday's report) source (WeChat, CN)
· Xinjian Transmission: a planetary roller screw manufacturer pushing for a ChiNext IPO, planning to raise RMB 2.822 billion to expand production; the company was the first to achieve large-scale mass production of planetary roller screws in China in 2022, and is a Tier-1 screw supplier for Tesla Optimus source (WeChat, CN)
· Jinghua New Material: its subsidiary Jingzhigan has established deep business collaboration with multiple dexterous-hand and embodied-intelligence companies source
· SK On: solid-state batteries for robots must first prove their economics — the battery maker struck a rare note of caution about this new customer segment, humanoid robots source
V. This Week's Observations
NHTSA to finish new driverless-vehicle rules by 2028: shifting from "what's in the car" to "what the car can do" · autonomy
NHTSA Administrator Jonathan Morrison confirmed to Bloomberg that the Trump administration plans to complete new autonomous-driving safety requirements by the end of this term, i.e. by 2028. The direction is to replace decades-old design-based rules that assume "a person sits behind the steering wheel" with performance-based standards: soliciting public and industry input to define what "behavioral capabilities" a self-driving car should have, then designing objective tests to determine compliance — for example, whether it can safely navigate around a double-parked fire truck, rather than what control hardware is installed in the vehicle. Morrison's framing centers on predictability: "manufacturers producing vehicles must be able to know for certain whether their vehicles meet the standard." Hardware-side deregulation is already underway: last month a proposal to amend FMVSS No. 135 to remove the brake pedal requirement for fully self-driving designs was put forward, mandatory steering wheels are also under review, and clauses on rearview mirrors, gear displays, wipers, sun visors, defrosting systems, and tire labels have all been proposed for removal — directly benefiting purpose-built robotaxis like Cybercab and Zoox. On the other side, regulatory pressure hasn't eased: Waymo, the largest robotaxi operator in the US (about 4,000 vehicles), recently recalled nearly its entire fleet and paused highway operations after vehicles drove through construction zones at highway speed; the NTSB and others have investigated multiple incidents of robotaxis passing stopped school buses; a January incident where a vehicle struck a child near an elementary school remains under investigation.Source: Automotive World source | Bloomberg source
Morgan Stanley: 2026 is the inflection year for robotaxis, a trillion-dollar market by 2040 · autonomy
Morgan Stanley's latest global robotaxi outlook judges the industry to be moving from technical trials to commercial scale-up, with a timeline of: mainstream operators reaching per-mile breakeven by 2028, scaled operations by 2030, a global fleet of 2.5 million vehicles by 2035, and a trillion-dollar-scale market by 2040. Cost deflation is the biggest catalyst — lidar and compute chip costs are falling 3-5% annually, and per-mile operating costs in some Chinese cities have already dropped below $0.50; the report cites two manufacturer-specific figures: Tesla's Cybercab targeting a cost of $0.30/mile, and Chinese vehicle BOM costs compressed to $35,000-40,000. Regionally, the report expects the US and China to lead (with Waymo and Tesla projected to split nearly 70% of US autonomous-driving miles by 2032), while Europe lags 3-5 years behind due to fragmented regulation slowing profit realization, and Japan and Korea treat it more as a solution to labor shortages. ⚠️ Single-bank view This diverges from Citi's assessment last week that "scaling up is a decade-long marathon" — the disagreement centers on the timing of breakeven.Source: Morgan Stanley research source (WeChat, CN)
TrendForce: companion humanoid robot output value only $1.1 billion by 2030 · humanoid
TrendForce forecasts that caregiving and emotional-companionship demand will push the companion humanoid robot market to $1.1 billion by 2030. The significance of this figure is that it puts a yardstick on the "humanoids pivoting to consumers" narrative — $1.1 billion is a rounding error next to Barclays' forecast last week that the humanoid market could reach $200 billion in a decade. This week UBTech was read as "sidestepping the labor-task logic, pivoting toward emotion and companionship, and going after consumers," but per this forecast, companionship alone won't support a leading manufacturer's core business, at least not before 2030.Source: TrendForce source
This week's supply chain roundup: harmonic reducer capacity has come online, and prices have collapsed · hardware
This is the most notable supply-chain signal this week, and it runs counter to the "supply chain chokepoint" narrative — it's not a shortage, it's a glut. On the capacity side: global harmonic reducer production capacity was 4.897 million units in 2024, jumping straight to 6 million units in 2025; against that, GGII data shows global industrial-robot harmonic reducer consumption was only 1.2187 million units in 2024 (795,500 in China) — existing capacity can absorb all global industrial-robot demand with plenty to spare. Prices have collapsed accordingly: the average price of a standard Chinese-made unit was about RMB 4,500 in 2023 and RMB 3,200 in 2024; by 2026, quotes of just over RMB 2,000 for bulk orders are common, with some smaller models falling below RMB 1,000 — a halving in three years. Three forces are compounding: after a single factory's annual output exceeds 300,000 units, economies of scale push unit cost down more than 30%; new entrants supply near cost price to secure test reports from benchmark customers (one company quoted 40% below the average price just to earn "designated supplier" status); and downstream demand has shifted from industrial robots to humanoids and collaborative robots, which are extremely cost-sensitive — cutting RMB 1,000 off a single reducer saves over RMB 10,000 on a full unit. But there's a hard floor to how far prices can fall: material and processing costs for a standard unit run about RMB 1,500; last year a manufacturer in southern China bought reducers priced at RMB 1,600, and within three months of deployment saw a 30% return rate. Leading makers are responding by digesting existing capacity rather than expanding — Leaderdrive's capacity utilization rose from 42.67% in 2024 to 67.76% in 2025, with output jumping from 251,700 to 433,700 units, while it pushed back a planned 1-million-unit capacity expansion project from the end of 2026 to the end of 2028; even so, its 2025 revenue of RMB 571 million rose 47% year-over-year, with net profit up 121% — prices are falling, yet profit growth is outrunning revenue growth.Source: Zhiti Xinju source (WeChat, CN) | Daoyi Institute source (WeChat, CN)
Reconciling demand: humanoid unit sales went from 50 to 15,000 in two years, reducers sell 430,000 units a year · hardware
Following on from the previous item, here's the demand-side accounting. Per MIR DATABANK, sales of complete-form humanoid robots in China were about 50 units in 2023 and about 15,000 units in 2025; a single standard humanoid uses more than 14 reducers, corresponding to third-party reducer sales of about 430,000 units for China's humanoid sector in 2025. The prevailing configuration is "harmonic drives for the upper body, planetary for the legs" — harmonic drives are dominated by Leaderdrive, Laifual, Radial Drive, and Tongchuan, while planetary drives are led by Zhongda Leader and NUODAR, which together account for over 70% of that market and primarily supply Zhiyuan Robotics and Unitree; cycloidal reducers sit between the two in load capacity and precision — theoretically a better fit for humanoids — but have complex manufacturing and higher costs, and remain in small-batch validation. Worth noting is a loosening at the component end: in the past only a handful of manufacturers worldwide could make 6mm micro harmonic drives, with overseas units selling for over RMB 2,000 each; Chinese manufacturers achieved mass-production breakthroughs this year and drove costs down to under RMB 1,000, recalculating the materials-cost math for dexterous hands. By contrast, another part of the chain remains tight: a humanoid uses 14 planetary roller screws, accounting for about 19% of a unit's total materials cost — more than harmonic drives and servo motors combined — and depends on imported high-end grinding machines with long validation cycles, which is why Xinjian Transmission is pushing for an IPO this week to raise RMB 2.822 billion for expansion. Reducer oversupply and roller-screw scarcity coexisting — that's the real shape of this supply chain right now.Source: MIR Rui Industry source (WeChat, CN) | Robot Chain source (WeChat, CN)
Top comments (0)