Today's Highlights

· Uber and Wayve launched the UK's first paid robotaxi service in London, with no lidar and no geofencing

· Figure signed a $3.5 billion compute deal with Nscale, with deployment of up to 100,000 GPUs

· According to Bloomberg, Waymo raised its first-ever debt financing of $3 billion, backed by Pimco, Blackstone and Sixth Street

· Horizon Robotics' licensing and services revenue surpassed hardware for the first time in H1, reaching 55% of total revenue

· PlusAI (Chinese autonomous trucking startup) attempted a SPAC merger for the third time, at a pre-money valuation of $800 million — a quarter of its 2021 valuation

· CJ Logistics put two humanoid robots to work on packaging in an Olive Young warehouse

· UBTECH's U1 received 13,361 pre-orders, with first deliveries on September 16; the VP called mass production a huge challenge

A single human demonstration teaches a dexterous hand to grasp 16 objects · manipulation

Data collection for dexterous hands has long been the bottleneck for multi-finger manipulation. A Stanford team (including Fei-Fei Li and Dorsa Sadigh) changed the reward from "task completion" to "whether the local geometry near contact points is correct." The resulting single policy transferred on real hardware across 16 objects, 4 task types, and 2 hand types, achieving a 71% success rate with the smallest sim2real gap among compared methods. Shape, size, mass, and friction can all vary as long as the contact structure is preserved.

Satvik Sharma et al. (Stanford) · arXiv 2609.01938 source

Before an emergency stop, a humanoid first calculates whether it can actually stop · locomotion

Executing a fixed motion upon receiving an emergency stop command is standard practice for current humanoids, regardless of whether stopping safely is actually feasible in the current state. Pieter Abbeel, Koushil Sreenath, Guanya Shi and others reformulated emergency stopping as a reach-avoid problem, paired with two complementary stoppability estimators: one supervised by the actual outcomes of a fixed stopping policy, and one supervised by physical recoverability derived via Hamilton-Jacobi reachability. Both only stop when stopping is deemed feasible; otherwise the system switches to a damped fall policy. The estimators don't depend on the behavior policy used before stopping, so they don't need retraining for new tasks.

Junfeng Long et al. (UC Berkeley / CMU) · arXiv 2609.02358 source

DeepMind releases a multi-view 3D point tracking benchmark; none of 30+ methods come close to solving it · benchmark

Multi-camera setups are increasingly used in robotics, AR/VR, and autonomous driving, yet no benchmark has tested long-horizon 3D point tracking with moving cameras and synchronized multi-view feeds. TAPVid-MV fills this gap: 284 sequences, 1,142 calibrated camera streams, and 109,769 point trajectories spanning robotics, human activity, driving, and synthetic scenes, with every trajectory manually visually verified. Across 30+ baselines, none come close to solving the task, and existing multi-view point trackers are not consistently better than monocular methods. By comparing reconstruction and tracking on the same data, the authors identify geometry recovery as the accuracy bottleneck.

Skanda Koppula et al. (Google DeepMind) · arXiv 2609.01899 source

Giving world action models depth perception · world-model

World Action Models (WAM) repurpose pretrained video diffusion models into robot policies that jointly predict future frames and actions — but they all consume only RGB, discarding 3D information. SA-WAM uses a nonlinear encoding to compress unbounded depth signals into the value range expected by a frozen VAE tokenizer, injecting geometry without modifying the tokenizer. A single diffusion backbone jointly outputs actions, RGB, and depth, achieving SOTA on RoboCasa and LIBERO-Plus, with even larger gains on a real UR5 in randomized environments.

Javier Alejandro Lopetegui Gonzalez et al. (Inria) · arXiv 2609.02531 source

Stepping stones and gap crossing: world models take over humanoid foothold decisions · locomotion

On terrain like stepping stones, narrow ladders, and gap crossings, feasible footholds are sparse and discontinuous, leaving no margin for a single misstep. WM-LOCO jointly trains a recurrent world model with a PPO policy, requiring no explicit foothold annotations, allowing the policy to plan beyond terrain that's immediately visible. In simulation, on both gap-crossing and stepping-stone tasks, baseline methods completely fail while this method succeeds; real-robot results have not yet been reported.

Yuxi Liu et al. · arXiv 2609.02542 source

Cross-embodiment zero-shot transfer — first, define what "zero-shot" actually means · vla

The term "zero-shot cross-embodiment" in VLA papers is often used inconsistently. ZETA splits it into two cases: the target embodiment never appears anywhere in the training data, versus one seen only during pretraining. With this distinction in place, the paper conducts a controlled analysis comparing four factors — including state-action representation — for their impact on tabletop manipulation transfer.

Mi Yan et al. · arXiv 2609.02546 source

Dissecting GR00T N1.5's action decoder · vla

VLAs connect language and perception to action, but what representations actually drive behavior internally has remained unclear. LAVLA proposes a latent-space clustering analysis framework, applying it layer by layer to NVIDIA's GR00T N1.5, with a focus on its action decoder.

Theodor Wulff et al. (Manchester / Hamburg) · arXiv 2609.02634 source

Human-to-humanoid motion retargeting, without hand-crafted mappings · locomotion

Turning large-scale human motion data into reference trajectories for robots is bottlenecked by differences in morphology, degrees of freedom, and joint ranges, with the standard approach being hand-written mappings for each robot model. UMR instead learns dense point-cloud correspondence, eliminating manually designed human-robot correspondence tables, with the goal of directly converting large-scale human motion libraries into trainable robot data.

Hanyang Cao et al. · arXiv 2609.02134 source

Figure | Strategic Partnership + Strategic Investment | $3.5 Billion Compute Commitment · humanoid

The compute order alone reaches $3.5 billion, with intended scale exceeding $6 billion. UK AI cloud provider Nscale signed a multi-year strategic partnership with Figure, with potential deployment of NVIDIA's Vera Rubin platform reaching up to 100,000 GPUs, with the first batch targeted for online in Barstow, Texas in the second half of 2027, supporting training and inference for Figure's Helix model series. Nscale also made a strategic investment in Figure, becoming both a shareholder and its preferred compute supplier. The agreement also leaves room for a reverse arrangement: the two companies will explore using humanoid robots to expand Nscale's own supply chain. A single compute commitment exceeding $3.5 billion from a humanoid robotics company is a scale previously seen only from foundation model companies.Source: Nscale source

PlusAI | SPAC Merger | Pre-money Equity Value of Approximately $800 Million · autonomy

Third time's the charm, after two prior failed attempts. PlusAI (Chinese autonomous trucking startup) reached a merger agreement with blank-check company Texas Ventures Acquisition III Corp, at a pre-money equity value of approximately $800 million. The deal could bring in up to $300 million in funding, including over $60 million in already-committed financing and about $236 million held in the SPAC's trust account. The company, headquartered in Santa Clara, California, focuses on SuperDrive, an L4 system for commercial trucks, targeting commercial deployment by 2027; its HyperFoundry software development platform has already generated $25 million in revenue, with full-year contracted revenue targeted at $40-50 million. It currently operates driverless freight routes in Texas with Ryder and International, and partners with global truck manufacturers on factory-installed production. Background: in May 2021, a merger agreement with HCIC V valued the company at about $3.3 billion, but was terminated that November; a second deal with Churchill Capital IX collapsed again in April 2026, officially attributed to "market conditions." Today's pre-money figure of $800 million is less than a quarter of the 2021 valuation.Source: Reuters source | 36Kr source

Waymo | First Debt Financing | Over $3 Billion · autonomy ⚠️ Unconfirmed report

Waymo, which has relied entirely on equity funding until now, is turning to debt markets. According to Bloomberg, the company is in talks with Pimco, Blackstone, and Sixth Street for its first debt financing of over $3 billion, with Goldman Sachs serving as arranger; the spread would be over 500 basis points above the benchmark rate and would carry no credit rating. Earlier this year the company raised $16 billion in equity at a $126 billion valuation. It currently completes over 500,000 paid rides per week across 14 US cities, with a year-end target of 20 cities and 1 million weekly rides.Source: Bloomberg source

UBTOTECH | Series A+ | Hundreds of Millions of RMB · humanoid

The round was led by Orient Securities Capital, with new investors Jinan Xianxing Investment Group and Zibo High-tech Zone Venture Capital participating, and angel-round investor Legend Star adding further investment. This Shandong-based company, founded in 2014, builds legged robots; its self-developed humanoid, "Walker Taishan" E1, won Shandong's delegation its first gold medal in the tug-of-war event at the 2nd World Humanoid Robot Games, and its quadruped products have already entered use cases including highways, telecommunications, power, tourism, and ports. Founder and chairman Fan Yong received his PhD from the Institute of Automation, Chinese Academy of Sciences in 2000, and the core team comes from Shandong University's Robotics Research Center and the Institute of Automation. The funding will go toward next-generation high-power-density integrated joint modules, industrial and consumer humanoid robots, and heavy-load, explosion-proof quadrupeds.Source: iThome source

First China humanoid robot ETF lists on Nasdaq · adjacent

Ticker CROB, issued by Defiance ETFs, tracks the Solactive China Humanoid Robotics Index, with constituents required to be headquartered in China and listed in Shanghai, Hong Kong, or Shenzhen. The top ten holdings — Green Harmonic Drive, Inovance Technology, Lingyi iTech, Sanhua Intelligent Controls, Megmeet, Hengli Hydraulics, Topgroup, Luxshare Precision, CATL, and Wolong Electric Drive — are exclusively component and manufacturing suppliers, with no full-robot makers included. Defiance CIO Sylvia Jablonski said China has already succeeded in EVs, solar, and batteries, and views humanoid robotics as the next stage of strategic positioning. Previously, Golden Rui Fund's KOID was the first US humanoid-themed ETF, which already includes Unitree Robotics.Source: China Fund News source

Uber and Wayve launch the UK's first paid robotaxi service in London · autonomy

Waymo has been beaten to the punch. Uber launched a robotaxi service in London, becoming the first company to offer paid autonomous rides in the city, using technology from UK startup Wayve. The fleet uses Ford Mustang Mach-E vehicles, with Nissan Leafs to follow; currently about 15 vehicles are registered with Transport for London. To ride, users must opt in via app settings, and there's no guarantee of a match when requesting UberX, Uber Comfort, or Uber Electric. Initially, a safety operator remains in the driver's seat, licensed by Transport for London; Uber says it does not intervene in driving and will phase this out. The technical approach diverges from the US mainstream in two respects: no lidar and no geofencing. Wayve describes its AV2.0 as a single learned AI driver that doesn't rely on hand-coded rule stacks or high-definition maps; spokesperson Julianne McGoldrick said the service will eventually cover all of London except airports. This marks the first large-scale test of driverless operation outside the US and China.Source: The Verge source | The Guardian source

Zoox brings paid robotaxis to Las Vegas airport · autonomy

Starting September 3, Zoox became the first company approved to operate driverless vehicles at Harry Reid International Airport, with pickup and drop-off available at both Terminal 1 and Terminal 3: at T1, drop-off is at the departures level and pickup at level 0; at T3, both pickup and drop-off are at level 0. The route connects the airport with the Las Vegas Strip. The company only began charging fares on August 10, under the first commercial exemption issued by NHTSA specifically for "purpose-built robotaxis." Local rideshare drivers have already publicly voiced concerns about their income.Source: TechCrunch | Las Vegas Review-Journal

CJ Logistics puts humanoid robots into real warehouse operations · industrial

Two humanoid robots have been deployed at an Olive Young distribution center in Yangji, Yongin, doing packaging work — placing cushioning material into boxes. The company says this marks the first time it has deployed humanoids in an actual logistics workflow, having previously only conducted on-site testing at its Gunpo fulfillment center. Plans call for expanding into picking, sorting, inspection, and packaging, with the goal of having a single robot handle multiple processes. On the technical side, CJ built its own robot foundation model (RFM), combining visual and sensor information with simulation data to let robots determine how to move around unfamiliar products and workstations; the hardware comes from Robotis, gripper technology from Aidin Robotics, and the RFM was developed by RealWorld AI. Two units don't amount to meaningful production capacity, but it shifts the question of whether humanoids can work in spaces designed for humans from demonstration to a production line shipping goods daily.Source: The Korea Times source

EACON's driverless system deployment surpasses 1,500 electric mining trucks · industrial

Nearly doubling in under six months, from 800 units in March to over 1,500 today. Within Easy Control's (Chinese mining autonomy company) autonomous driving deployments, pure-electric mining trucks now account for about 42% of the fleet, hybrid diesel-electric 41%, methanol hybrid 16%, and conventional internal-combustion trucks under 1%. At China Energy's Zhundong open-pit coal mine, 120 pure-electric trucks run on the ORCASTRA system with 28 charging stations, with dispatch simultaneously calculating battery levels, predicting energy consumption, tracking charging-station availability, and coordinating with production schedules — vehicles navigate to charging stations on their own and return to the transport line once charged. At Zijin Mining's Julong copper mine, 60 Sany SKT105ED trucks operate routinely at high altitude, maintaining an operational rate above 90% under low air pressure, extreme cold, and freeze-thaw conditions with rain and snow. China Baowu's Taihe iron mine added 12 more trucks in March after over a year running its initial 6 units. Mining sites are currently one of the few areas in commercial driverless vehicles where electrification and autonomy have both been achieved at scale together.Source: International Mining source

UBTECH's U1 delivery test approaches, as the substance behind 13,361 pre-orders remains to be seen · humanoid

First deliveries are set for September 16, with the company aiming to fulfill all orders within the year. The U1, launched on June 30, is a hyper-realistic companion robot aimed at emotional companionship; the launch event announced cumulative pre-orders of 13,361 units across all channels. That figure comes with caveats: pre-orders require only a RMB 3,000 deposit, which is fully refundable during the presale period, leaving the eventual final-payment conversion rate unknown. Production capacity is another hurdle — as of the end of 2025, UBTECH Robotics' annualized production capacity for full-sized humanoids was about 6,000 units, while the U1 is an entirely new consumer model requiring dedicated production lines for its silicone skin and facial micro-servos. Vice President Jiao Jichao told reporters after the launch event bluntly: "Mass production at a scale of over ten thousand units is a huge challenge." He attributed the difficulty to low yield rates on lifelike silicone skin and supply chain constraints from the large number of non-standard micro-servos throughout the robot. JD.com's flagship store lists the U1 Pro as available after September 15, with customer service quoting a 60-day shipping window after ordering. Capital markets reacted with initial excitement followed by a pullback: on launch day the stock rose over 18% intraday before closing up 7.48% at HK$102.8; by the September 3 close it had fallen back to around HK$79, a pullback of over 20% from the peak. On August 26, the U1 won the "AI-Driven Healthcare Innovation Award" for Asia-Pacific at the MedTech World Awards 2026 in Hong Kong; around the same time, UBTECH also jointly launched a wellness-care edition of its Umi robot with Meditech.Source: Red Star Capital Bureau source

Zhuji Dynamics' TRON 2 ties rebar on a construction site · embodied

Two TRON 2 dual-arm robots, one holding a nail gun and the other a wooden board, worked together to assemble and secure formwork, then laid and tied rebar across multi-layer scaffolding, completing a scaled-down construction process for a concrete tilt-up panel. The work was carried out by physical AI company ZINOVA, in partnership with construction robotics company RIC Robotics. ZINOVA's approach centers on "tool intelligence" — building its technology stack around general-purpose tool grasping and tool-interaction perception, aiming to make the same system reusable across tools, tasks, and robot embodiments rather than customizing a robot for every trade. Construction was chosen as its first validation scenario, citing both a shortage of skilled labor and the highly non-standardized nature of construction environments. The TRON 2 platform supports switching among dual-arm, biped, and wheeled-biped configurations, and was reconfigured here for large-space, multi-process requirements. This remains a scaled-down proof of concept for now, not routine operation on an active construction site.Source: Gasgoo source | 21st Century Business Herald source

Horizon Robotics' licensing revenue surpasses hardware for the first time in H1 · autonomy

The revenue mix has flipped. Horizon Robotics' H1 revenue reached RMB 2.055 billion, up 32.9% year-over-year, of which licensing and services revenue was RMB 1.129 billion, up 52.7% year-over-year and accounting for 55% of the total — surpassing, for the first time, the RMB 926 million from products and solutions (up 14.8% year-over-year). Overall gross margin held at 66%, with gross profit of RMB 1.356 billion. On market share, the company's share of the Chinese-brand ADAS market surpassed 50.0% for the first time; its share of the urban NOA chip market rose from 17.9% in 2025 to 22.8%, moving from third to second place; and it holds the top spot in the overall Chinese-brand smart driving chip market with 31.9% share. Capital markets have long viewed Horizon Robotics primarily as an automotive chip company, but this half-year report shows the higher-margin licensing business is now the larger half.Source: Caixin source

Altman confirms for the first time: OpenAI will build its own humanoid robot · humanoid ⚠️ Remarks in a podcast

In a podcast released local time on September 2, Sam Altman, when asked about hardware progress, said "we will definitely do humanoid robots." His reasoning: the physical world is built around the human body — keyboards, doors, heavy machinery, kitchens. He also said OpenAI will simultaneously develop specialized robots for data centers, with near-term focus remaining on industrial infrastructure and data center automation, as well as the underlying cognitive models. No specific model, timeline, or team size was given.Source: The Paper source

Beijing's humanoid data facility opens; RoboMIND downloads surpass 20 million · world-model

The Embodied Intelligence Data and Training Base of the Beijing Innovation Center for Humanoid Robotics opened today. Built in under a year, the facility spans more than 6,000 square meters, covering six major categories — home, retail, office, industrial, medical, and wellness — across more than 30 representative scenarios, with over 150 types of robot hardware and more than 100 embodiment-free data collection setups. It has an annual capacity of 180,000 hours of high-quality data and has already delivered over 30,000 hours to leading clients. Its open-source dataset RoboMIND doubled in global downloads within a month, surpassing 20 million. Facility director Xia Hualin said the most critical competitive edge in embodied intelligence going forward won't be the robot body or the algorithm, but the data flywheel: "I originally came from autonomous driving. The biggest factor behind how quickly autonomous driving was deployed wasn't the vehicle or the large-model algorithms — those were already mature. What was needed most was data." He defined the most valuable data as "feedback data" — for example, if a robot can wash dishes at Household A but fails at Household B, quickly feeding the failure data from Household B back into training and redeployment. The facility says AI-model-driven automated labeling has improved data production efficiency by over 50% and reduced overall costs by over 30%; in H1, the data acceptance pass rate for a leading client delivery exceeded 95%.Source: Sina Technology source

LG Electronics teams up with NC AI to take on South Korea's edge-AI chip humanoid track · humanoid

South Korea's Ministry of Trade, Industry and Energy, together with the Korea Evaluation Institute of Industrial Technology (KEIT), selected a consortium of NC AI and LG Electronics for the humanoid track of its "K-Edge AI Semiconductor Technology Development Project." The goal is to use domestically made edge-AI chips to power autonomous humanoids capable of performing a variety of tasks in complex everyday environments, bringing robot hardware, training data, AI models, and domestic chips together under a single development framework. NC AI's (Korean gaming company's AI unit) role covers initial training data for General Motion Tracking (GMT) and a human-to-robot motion retargeting pipeline; it plans to use its own motion capture and animation production technology, combined with digital twin environments, to mass-generate task scenarios rather than collecting them one by one in real-world settings.Source: The Elec source

Unitree's wholly owned subsidiary lands in Chengdu, with a "three bases, two centers" layout · humanoid

Chengdu Yuchen Technology Co., Ltd. completed its business registration on September 2. Unitree Robotics' Chengdu plan calls for three bases plus two centers, including a planned flagship global launch center on Chunxi Road. As previously reported, Unitree's market cap had halved from its first trading day by its 11th trading day since listing; new this week is this regional expansion, along with internal management controversy sparked by the company's employee expense-reimbursement policy.Source: Pandaily source | Sina Finance source

Tesla's Cybercab event set for Thursday, with very few details · autonomy ⚠️ Planned event, unconfirmed details

Reuters reports that Tesla will hold a Cybercab-related event on Thursday, with the company not disclosing specific content. Around the same time, reports have surfaced of Austin residents already riding in Cybercabs. Production status, capacity, and the scope of any regulatory permit remain without verifiable public figures.Source: Reuters source

France launches Tesla FSD road testing, just two months after calling it "too risky" · autonomy

On September 2, Paris approved on-road testing of Tesla FSD in France, with Tesla providing two vehicles equipped with the system. French Transport Minister Philippe Tabarot said France would independently verify data provided by the Netherlands and Tesla, and test the system under French road conditions, with completion planned for mid-to-late September. This marks a clear reversal — in July, he had said the safety tradeoffs weren't yet sufficient to justify authorization, but he now describes "constructive exchanges" with Elon Musk, with the two sides having cooperated on technical adaptation for months. An EU-level approval vote is imminent, and France's test findings will feed directly into that process.Source: Euronews | Reuters

China's first systematic overhaul of its road traffic law gives autonomous driving its own dedicated chapter · autonomy

The draft revision comprises 9 chapters and 170 articles, submitted for first reading at the 24th session of the Standing Committee of the 14th National People's Congress on August 25, and opened for public comment starting August 28; within less than a week, over 8,000 people had submitted over 12,000 comments. The current law has been in effect since 2004, with only three prior amendments touching individual provisions. The new dedicated chapter on "Special Provisions for Autonomous Vehicles" establishes several firm rules: autonomous vehicles must pass a road traffic rules compliance test and be legally registered before operating on public roads; in the event of a traffic violation while autonomous driving functions are activated, the manufacturer or importer is held responsible, and if the company believes the violation is unrelated to the autonomous driving function, the burden of proof falls on the company itself; and traffic police authorities may retrieve relevant vehicle data in the event of an accident. NPC Standing Committee member Huang Ming said the purpose of the dedicated chapter is to "draw clear regulatory red lines for the new business model and clarify a compliant development path." The reversal of the burden of proof — previously accident attribution had long relied on manufacturers' own claims — has now been written into the draft law.Source: China News Service source

Goldman Sachs sharply raises humanoid shipment forecast: 890,000 units by 2030 · humanoid

In a research note published September 2, Goldman Sachs raised its forecast for global humanoid robot shipments in 2030 from a previous 256,000 units to about 890,000 units, and its 2035 forecast from 1.4 million to about 6.5 million units, corresponding to a market size of about $138 billion; this year's baseline shipment estimate was also raised from 51,000 to 75,000 units. On deployment sequencing, the bank ranks logistics and warehouse operations as the leading use case, with the automotive industry as the second wave. The nearly 3.5x upward revision compares against the same institution's 2035 market-size forecast of $38 billion issued just over a year ago.Source: China Review News

Arthur D. Little: humanoids account for only $2-3 billion of an $18 billion physical AI market · adjacent

A Blue Shift report titled "I, Robot," published September 3, offers a contrasting figure to the item above: the current overall physical AI market is about $18 billion, with humanoids accounting for only $2-3 billion of that; by 2030, the overall market could reach $60-100 billion, growing at over 20% annually. The report's assessment is that, in the near-to-medium term, value is captured mainly by specialized physical AI systems targeting specific high-value tasks, with humanoids remaining relatively immature by comparison. Blue Shift director Albert Meige put it this way: "When it comes to physical AI, humanoids are the moonshot that captures attention, capital, and talent, but real-world value is lagging behind the hype." In the same week, an investment bank raised its shipment forecast more than threefold, while a consultancy said humanoids account for only about one-sixth of the current market — the two figures aren't directly in conflict, since one measures future shipments and the other measures current revenue.Source: Arthur D. Little | Intelligent CIO source

This week in supply chain: Chinese sensor makers' market concentration solidifies, while reducer makers expand capacity and cut prices simultaneously · hardware

Consolidation has already taken hold in force sensing and end-effectors. The "2026 China Humanoid Robot Six-Axis Force Sensor Market Research Report," published by CCID and China Electronics News on August 31, shows that the top three Chinese manufacturers together hold 92.8% of the market, with Bluepoint Sensing above 80%, Kunwei Technology at 9.2%, and Sunrise Instruments at 3.3%; per the report's baseline year, Chinese shipments of six-axis force sensors for humanoid robots surpassed 10,000 units for the first time, with market size entering the hundreds-of-millions-of-RMB range. The absolute volume remains small, but domestic substitution is already essentially complete, with the next variable being unit price rather than market share. Tactile sensing is ramping up in parallel: on August 31, Folisenol Technology announced it has cumulatively delivered over 30,000 sets of tactile sensors to Lingxin Qiaoshou (Chinese dexterous-hand maker), part of a 100,000-unit framework order the two companies signed in February.

Actuators and transmission components are on a different trajectory. Green Harmonic Drive's H1 revenue was RMB 349 million, up 38.64% year-over-year, with net profit attributable to shareholders of RMB 70.11 million, up 31.25% year-over-year; the company previously said its harmonic reducer shipments reached nearly 500,000 units last year, with this year's capacity being prepared for close to a million units. In the same period, Zhongda Leader's (Chinese motion-control component maker) net profit attributable to shareholders fell 40.09% year-over-year (previously reported). Within the same downstream market and the same quarter, leading and second-tier players are moving in opposite profit directions, with price-war pressure not yet reaching the company with the largest market share. Taken together, the tightest constraint in the humanoid supply chain this week isn't capacity but unit pricing: sensors, having just crossed the 10,000-unit mark, are already dominated by Chinese suppliers, while reducer makers have built out capacity ahead of demand materializing.Source: Cailian Press | Jiemian News | Shanghai Securities News