Today's Highlights
· Waymo's Mesa factory inventory shows 953 Robotaxis, with 684 already converted and ready for deployment
· Uber partners with Hinomaru Kotsu to pilot Robotaxis in Tokyo by year-end
· Deep Sea Smart Human (Chinese deep-sea robotics startup) raises over RMB 500 million in Series A, with TotalEnergies' fund joining alongside state-owned capital
· UOTECH (Chinese servo joint maker) raises over RMB 200 million in Series B, with 1.5 million joint units in the order backlog
· NEURA acquires Bosch Rexroth's ACTIVE Shuttle autonomous transport business
· BYD's 31-DoF humanoid "Xiaodi" starts working as a greeter at dealerships
· G0.5 unifies reasoning and action in a single autoregressive flow, achieving 76.7% on real robots, surpassing π0.5
I. Research Progress
G0.5: A Single Autoregressive Flow Outputs Reasoning and Action Together · vla
The mainstream VLA recipe today is "pretrained VLM + separately trained flow-matching action expert," which degrades the VLM into a mere context encoder rather than a decision-maker. G0.5 instead has a single transformer decoder output both reasoning tokens and action tokens under one objective, scaling this approach to foundation-model size via a learnable cross-embodiment action tokenizer, a native chain-of-thought flow, and a multi-second visual memory module. Because reasoning and action share the same weights, the VLM's instruction-following ability transfers directly to physical behavior, and prompts can adjust action granularity and task duration without retraining. It beats SOTA on 7 independent benchmarks: 76.7% on real-robot fine-tuning with R1lite/R1pro (vs. 53.3% for π0.5 and 24.4% for GR00T-N1.7), 31.4% on the 50 long-horizon household tasks of the 2025 BEHAVIOR Challenge (vs. the challenge-winning 26.1%), 98.9% on LIBERO, and 93.3% on RoboTwin 2.0.
Yicheng Liu et al. · arXiv 2608.11739 source
StellaVLA: Give the Policy a "Structured Demonstration" and It Learns New Scenes at Test Time · vla
VLAs tend to break down when the scene, viewpoint, or objects change, usually requiring re-collected data and fine-tuning. StellaVLA instead retrieves just one demonstration at test time and injects it as context — and this demonstration is automatically converted by an offline pipeline into a task plan, sub-goal descriptions, and verbalized 3D motion, conveying "why to do it this way" rather than pixel trajectories, which makes it reusable across embodiments (real robots, human hands, XR demonstrations) while adding no inference-time latency since only the action expert runs at inference. It topped the August 1 VLA-Arena leaderboard with a total score of 0.63, versus 0.44 for π0.5 and 0.22 for LingBot-VLA, with a 98.8% average success rate on LIBERO.
Siyu Xu et al. · arXiv 2608.11671 source
RIFT: World-Action Models Can "Imagine" the Future Without Rollout · world-model
World-action models (WAMs) decide actions by predicting future frames, but generating video frame-by-frame at deployment drives up latency. The authors ran closed-loop interventions on 4 WAMs across 40 LIBERO tasks and found that replaying a fixed future K/V cache for Joint and Cosmos-2 barely affects execution, showing that "consuming the cache" and "producing the cache" can be decoupled. RIFT therefore uses a learnable anticipation token to construct the complete future cache in a single forward pass, matching the rollout approach's 98.8% LIBERO success rate while cutting action-chunk latency by 68.2%–89.1%.
Chushan Zhang et al. · arXiv 2608.11521 source
ForeWAM: Feeding Dynamics to the Action Head Without Decoding Future Video · world-model
Another rollout-free approach appearing the same day as RIFT: explicit future-frame WAMs require iterative denoising of video, which is expensive, while direct-policy WAMs are fast but expose no interface for predicted dynamics. ForeWAM performs only a single Video DiT prefill via Future-KV, and supervises a "dynamics register" with a frozen latent-action teacher so that the implicit future state captures object motion, contact changes, and task progress; real future frames are only needed during training, and deployment never generates future frames at all. Without pretraining on embodied robot data, it achieves 96.7%/96.9% average success on LIBERO.
Jiakai Huang et al. · arXiv 2608.11605 source
MiDAS: One Demonstration Plus Autonomous Interaction Teaches a General Policy New Tasks · vla
Fully autonomous learning is stuck on sparse rewards and weak zero-shot exploration — robots can hardly stumble into successful behaviors from scratch. MiDAS instead anchors a pretrained VLA to a target task with just one (or very few) demonstrations, then runs value-based online RL on a residual policy. On LIBERO and RoboCasa, strong performance is recovered from a single demonstration; on a dual-arm YAM real robot, a fragile policy learned from one demonstration developed new successful behaviors after roughly 6 hours of online interaction — the authors describe this as the first reliable adaptation of a robot policy from a single task demonstration.
Shreyas Kowshik, Max Simchowitz et al. · arXiv 2608.11363 source
SHAPER: Self-Upgrading "Skills + Framework" Through Evolution, Without Touching a Single Parameter · vla
HF trending +8. When fine-tuning is too costly or the robot API isn't programmable, how else can an embodied agent get stronger? SHAPER keeps model parameters frozen, using the same model as both planner and optimizer, repeatedly evolving reusable skills and a context-code harness through rollouts in the target environment. Comparisons against pure execution, supervised fine-tuning, and test-time-scaling baselines on VLABench and ESI-Bench show that "skill and framework optimization" is a viable path when training isn't available.
Peidong Wang et al. · arXiv 2608.11350 source
Handedness Bias in Bimanual Humanoids: Initial Pose Quietly Decides Success · manipulation
Aggregate success rates can mask pose-related failures. The authors quantify VLA-induced "handedness priors" across multiple policies and 17 initial configurations, finding that success rates for the same pose vary widely across policies, that certain initial arm configurations induce or suppress asymmetric hand preferences, and that the presence or absence of a wrist camera also affects the ranking of policies and their scores; expanding initial-pose coverage in the training set and applying targeted augmentation to low-scoring configurations both markedly improve robustness.
Chaeyeon Jung, Juyoun Park · arXiv 2608.11769 source
RoboSaGA: Saliency-Guided Augmentation to Withstand Shadows and Background Clutter · manipulation
Already cited 11 times. Standard random-crop and color-jitter augmentations used in visual behavior cloning fail under large visual domain shifts; RoboSaGA uses saliency-guided augmentation to maintain in-domain performance while significantly improving robustness to distractors, background changes, and lighting/shadow variation.
Zheyu Zhuang et al. · arXiv 2608.11870 source
Other papers today: SMPC demonstrations plus sparse offline-to-online RL for loco-manipulation, avoiding dense reward shaping (arXiv 2608.12063 source); HandEdit, the first unified benchmark for egocentric "human hand → robot hand" image editing (arXiv 2608.12122 source); DaViNCi, the first outdoor vision-language navigation dataset with both continuous action and dynamic elements (arXiv 2608.11901 source); Map-Det3D, single-view multi-view detection using feed-forward metric 3D reconstruction priors (arXiv 2608.12179 source); RoadWeaver, generating large-scale lane-level HD maps from scratch for autonomous-driving simulation (arXiv 2608.11580 source); TrafficDiffuser, traffic scene generation via joint diffusion of initial and goal states followed by trajectory completion (arXiv 2608.11407 source); Video2Track, converting real interaction videos into adversarial tests with tunable risk in closed venues (arXiv 2608.11592 source); linearly decoding early-warning signals of imminent failure from OpenVLA's internal activations (arXiv 2606.29699 source); HUGIN, joint vision-language planning across multiple scenarios for logistics sorting (arXiv 2608.11692 source); Thea, an embodied harness that packages robot capabilities as tools orchestrated by an agentic loop (arXiv 2608.11246 source).
II. Funding & Deals
Deep Sea Smart Human (Chinese deep-sea robotics startup) | Series A | Over RMB 500 million · industrial
New investors include Guanghe Venture Capital, Fortune Capital, GGV Capital, Yida Capital, China Insurance Investment, and CETC Investment, along with an energy technology fund managed by Cathay Capital with TotalEnergies as the funding entity; existing shareholders including Zhengxuan Investment, Yunze Capital, Yunshi Capital, and Gaojie Capital made large follow-on investments. Founded in 2023, the company builds work-class ROVs operating at depths of several thousand meters — heavy equipment essential to deep-sea oil and gas, subsea communications, offshore wind, and deep-sea mining, a market long dominated by European and American manufacturers. According to data obtained by ChinaVenture, as of this July the company had signed several hundred million RMB in new deep-sea robot equipment sales contracts, with a similarly-sized order backlog locked in for the next two months and another RMB 1 billion in potential orders under negotiation. This marks the first time four market-driven top-tier VCs have invested in deep-sea robotics simultaneously, signaling embodied AI's spillover from land to sea. Source: ChinaVenture source
UOTECH (Chinese servo joint maker) | Series B | Over RMB 200 million · hardware
12 institutions participated including Samsung Venture Investment, Guofang Venture Capital, and Jiaqiao Capital, with existing shareholders Jiaqiao Capital, Fudar Holding, and Xishan State-owned Investment continuing to add on; the company has completed three funding rounds this year alone, raising over RMB 500 million in total. This integrated servo-joint manufacturer says its joint order backlog for 2026 has already exceeded 1.5 million units, 15 times its total 2025 shipment volume; its fully automated production line built in Shanghai's Zhangjiang area earlier this year produces one harmonic joint every 60 seconds with a 98% first-pass yield, and its Wuxi base has purchased an additional 34 mu (about 5.6 acres) of land for a new automated facility. Its client list includes AgiBot (Chinese humanoid robotics company), iFlytek, FAW, Midea, and Siemens. Joints are a key bottleneck for humanoid whole-machine cost and consistency, and the jump in orders from "prototype supply" to million-unit scale is a rare hard data point of certainty in the supply chain. ⚠️ Manufacturer's figuresSource: China.com source
NOUSBOT | Series A+ | Nearly RMB 100 million · hardware
Led by Shunwei Capital with participation from Wangqian Capital, the funding will mainly go toward expanding production of miniature planetary roller screws, product iteration, and overseas expansion. Founded in July 2023, the company grew out of a family business with over 30 years of screw-manufacturing experience, and its shareholder history includes SAIC Capital, Shanghai Semiconductor Industry Investment, and Legend Capital. Founder Xu Yang believes screws will be the ultimate solution for humanoid linear joints — rotary joints are easy to control but struggle to combine high force with a compact size, and stall-induced heating remains unresolved, whereas a screw's self-locking force lets the motor operate intermittently. He also noted a shift in requirements: dexterous hands used to be "nearly unusable," frequently failing after about 200 hours, but transmission demand has now extended from hands to wrists, forearms, and even the neck and face, with neck degrees of freedom rising from one or two to four to six.
Source: 36Kr source
NEURA Robotics Acquires Bosch Rexroth's ACTIVE Shuttle Autonomous Transport Business · industrial
The deal takes effect October 1, with NEURA Mobile Robots taking over ACTIVE Shuttle's hardware, software, and ongoing customer service. This autonomous transport system is already used by numerous industrial companies for assembly-station replenishment and small-load transfers, a proven, field-tested product line, and NEURA plans to fold it into its own Physical AI ecosystem while adding AI capabilities and interfaces. Founder David Reger's stated goal is a shared technology base across mobile robots, industrial robots, and humanoids, rather than separate intelligence stacks for each robot category. The German company just closed a Series C of up to $1.4 billion in June, and this acquisition of a mature business from Bosch is one example of European robotics assets consolidating around new market entrants. Source: Tech.eu source
Anthropic | In Talks to Acquire Decart AI | ~$6 billion · world-model
Reuters, citing sources, reports that Anthropic is in talks to acquire Israeli AI startup Decart at a valuation of around $6 billion, targeting its video generation and chip optimization capabilities. Decart has been known for real-time video generation and world models; if the deal closes, it would signal that foundation-model companies are starting to treat "real-time generative worlds" as a must-have capability rather than something to source externally. ⚠️ Rumored termsSource: Reuters source
Hai Robotics | Hong Kong IPO Application Lapses · industrial
The Hong Kong listing application status for this box-based automated storage (ACR) leader has changed to lapsed; the company had previously disclosed a global ACR solutions market share of over 30%. Amid a wave of logistics robotics companies filing for listing, this is a rare process interruption. Source: FX168 Finance source
Unitree Robotics | STAR Market IPO | Share Payment Completed · humanoid
Following an online subscription that was oversubscribed 8,288 times, payment for this offering has been completed; online investors forfeited subscriptions for 8,734 shares, which will be underwritten by the lead underwriter. On the institutional placement side, 89 mutual funds collectively received allocations worth nearly RMB 1.8 billion, and 37 insurers took roughly 30% of the institutional tranche. Reports of scalpers' resale prices reaching RMB 410 trended on social media, though some experts said this reflects sentiment rather than fundamentals. Source: Sina Finance source
Wanji Yizu | New Funding Round | JD.com Enters Robot Leasing · adjacent
Following an earlier strategic round valuing it at RMB 1 billion, this robot leasing platform has secured a new round of funding, with JD.com simultaneously entering the sector. Leasing is a common way to sidestep the barrier of one-off customer purchases, and platform-level funding often runs ahead of the actual scaled delivery of the equipment itself. Source: Sina Finance source
III. Commercialization & Deployment
AgiBot's Dual-Arm Robot Live-Sorts 1,816 Items/Hour, Beating the Figure 03 Benchmark by 45% · embodied
During a continuous one-hour livestream, a dual-arm robot equipped with only a standard industrial gripper completed 1,816 valid sorts, exceeding the 1,248 items/hour benchmark set by Figure 03's earlier 200-hour continuous run by 45% in short-window throughput terms; according to public information, this solution cuts costs by 70% with over 98% sorting accuracy. Behind it is AgiBot's WALL-B, released in April, which uses a "world unified model" architecture that jointly trains vision, language, action, and physical-change prediction in a single network. It should be noted that livestream throughput is a capability demonstration rather than a mass-production-line metric. ⚠️ Manufacturer's figuresSource: iFanr source
Waymo's Mesa Factory Inventory Shows 953 Robotaxis, 684 Conversions Completed · autonomy
According to a monthly aerial survey cited by Electrek, Waymo's factory in Mesa, Arizona currently holds 953 dedicated Robotaxis, of which 684 — about 72% — have completed conversion and are deployment-ready, with 269 still awaiting conversion; the same count stood at roughly 500 in early June. Also confirmed for the first time were 12 complete vehicles based on the Hyundai Ioniq 5 platform, marking the first appearance of Waymo's second dedicated vehicle model on the completed list. The Ojai model's chassis is produced by Zeekr (Chinese EV brand under Geely) before being shipped to Mesa, where Magna installs the sixth-generation autonomous-driving system — 13 cameras, 4 lidars, and 6 radars, a 42% reduction in sensor count from the previous generation, with Waymo saying hardware cost can be kept under $20,000 per vehicle. Its weekly paid rides already exceed 500,000, with a target of 1 million within the year, while its operating fleet stood at roughly 2,500–3,000 vehicles at the start of the year — meaning the converted, completed vehicles already represent a substantial pool of pending capacity, and the next variable shifts from "can it be built" to "how fast can it be deployed." ⚠️ Estimated figuresSource: Digital Today (citing Electrek and aerial survey data) source
Uber Partners with Hinomaru Kotsu to Pilot Robotaxis in Tokyo by Year-End · autonomy
Uber announced a partnership with veteran Tokyo taxi operator Hinomaru Kotsu to deploy a Robotaxi pilot fleet, expected to launch by the end of 2026. Uber's autonomous-driving footprint in Japan had previously centered on Tokyo partnerships with Wayve and Nissan; this time a local fleet operator takes on the operational side — a typical division of labor for overseas Robotaxi deployment, with the technology provider supplying the system and the licensed fleet operator supplying capacity and compliance. Source: 36Kr source
BYD's Humanoid "Xiaodi" Enters Dealerships, 31 Degrees of Freedom · humanoid
"Xiaodi" made its debut at BYD Space experience centers in Zhengzhou, Shenzhen, and Shanghai — 1.61 meters tall, weighing 58 kg, with 31 degrees of freedom (7 in the hands alone), supporting six Chinese dialects and six foreign languages, with facial recognition, gesture recognition, and lip-reading capabilities, handling greeting, vehicle introduction, and Q&A duties. BYD says it plans to deploy two to three units per dealership. This puts a humanoid into an actual retail setting rather than a trade-show booth, though the role remains a sales-greeter type of task rather than production-line work; around the same time, Xpeng plans to bring its Iron humanoid into Chinese dealerships by Q1 2027. One practical reason automakers are choosing retail over factories for this rollout is that they already have an extremely wide dealership network they can use as a large-scale real-world test bed. Source: Interesting Engineering source
EACON Begins Unmanned Night-Shift Operations at Australia's Havana Mine · autonomy
EACON (Chinese mining autonomy company) has launched night-shift autonomous haulage operations at Australia's Havana mine. Night shifts are one of the truest tests of mine-site autonomy — poor lighting, high labor costs, and elevated accident rates mean that consistently running a night shift usually signals the system has moved past pure daytime pilot operation. Source: International Mining source
WeRide Considers Bringing Robotaxis to Australia and More Asian Markets · autonomy
The company is evaluating expansion into Australia and Asian markets, having already seen its overseas revenue growth clearly outpace its Chinese revenue growth. This remains at the market-evaluation stage, with no licensing or launch timeline yet confirmed. ⚠️ Planning stageSource: Capital Brief source
IV. Industry Developments
2nd World Humanoid Robot Games Open August 22, Adding 14 New Real-World Scenario Events Across 1,301 Matches · humanoid
This edition features 51 events across 1,301 matches, with 666 teams from 16 countries participating — a substantial expansion over the first Games — and includes night sessions for the first time. The most notable change is in the format: 14 new scenario-based events shift competition from gymnasium running-and-jumping toward "last-mile" real-world settings — pick-and-place, delivery, and human-robot collaboration in settings like hotels, retail stores, and production lines. A 5-on-5 soccer match has already kicked off. The Games are primarily a showcase, but the addition of scenario events adds a further function: comparing different manufacturers' actual task completion under a unified rule set on similar tasks — something ordinary manufacturer demos, each on its own terms, can't provide. Source: The Paper source
JD Logistics Half-Year Report: AI and Automation Scaled Across End-to-End Supply Chain, R&D Spending Up 53% · industrial
JD Logistics disclosed that AI and automation technologies have been deployed at scale across end-to-end supply-chain scenarios in the first half of the year, with group-level R&D spending up 53% year over year, positioning JoyAI externally as its "physical-world intelligence foundation." E-commerce logistics is one of the earliest sectors in China to create stable, paid deployment slots for robots, and its capital expenditure direction typically leads general embodied AI deployment by one to two steps. Source: Futu Bull source
Tokyo Electron Expands Partnership with Nvidia Toward Agentic AI and Robotics · world-model
Semiconductor equipment giant Tokyo Electron is expanding its partnership with Nvidia to cover agentic AI and robotics. Nvidia has been signing a string of deals with Japanese industrial players recently, following the same playbook as in automotive and healthcare — first embed simulation and foundation models into a partner's existing production lines, then move on to the robot itself. Source: Evertiq source
Daedong Partners with Rainbow Robotics to Develop Agricultural Humanoid "AGROID" · humanoid
South Korean farm-machinery giant Daedong and Samsung-affiliated robotics company Rainbow Robotics announced joint development of an agricultural humanoid robot called AGROID. Agriculture is one of the few settings where "labor shortage" and "unstructured work" both hold simultaneously, which is why it's often seen as a way for humanoids to sidestep the strict cycle-time demands of factories.
Source: Asia Economy source
Honda Unveils New Humanoid "Avatar" · humanoid
Honda has unveiled a new humanoid robot, Avatar, widely seen as a continuation of the ASIMO lineage. Japanese OEMs have kept a low profile through this humanoid boom, so Honda's re-emergence is itself a signal. Source: Robotics & Automation News source
UBTECH Partners with Maoming to Build Embodied AI Application Innovation Center · humanoid
The project has cleared filing, with planned investment of RMB 100 million. Local governments trading scenarios and funding for OEM deployments has become one of the most common partnership formats in humanoid robotics this year, though actual delivery volumes usually aren't clear until acceptance data comes in one to two quarters later. Source: Sina Finance source
Geek+ Launches RoboShuttle Hyper Series Autonomous Mobile Robots · industrial
A next-generation AMR product line for warehousing, emphasizing higher-throughput box handling. Source: GlobeNewswire source
Hardware & Supply Chain
· Zhenghe Industrial (Chinese chain-drive maker): This chain-drive leader's roughly RMB 700 million private placement was approved, with proceeds directed toward high-end farm-machinery alloy cutting tools and robot dexterous-hand components — another example of a traditional transmission-parts manufacturer moving into the embodied AI supply chain source
· Texas Instruments CAN XL Transceiver: A new-generation transceiver for higher-bandwidth automotive and industrial buses has launched; the bandwidth demands of multi-joint, high-frequency force control in robots are moving up the same curve source
· Wieson Technologies: First-half revenue of NT$1.567 billion, with second-half focus shifting toward AI thermal management and robotics-related business source
V. This Week's Watch
China Issues Mandatory National Standard for Autonomous Driving System Safety, First Unified Access Threshold for L3/L4 · autonomy
The "Intelligent and Connected Vehicles — Autonomous Driving System Safety Requirements" (GB 44721—2026), developed under China's Ministry of Industry and Information Technology, has been approved for release by the State Administration for Market Regulation and the Standardization Administration, and will take effect on July 1, 2027. It applies to M- and N-category vehicles equipped with L3 or L4 autonomous driving systems, but excludes automated parking systems. The key change is legal force: its predecessor, GB/T 44721—2024, was a recommended standard that companies "could" follow; as a mandatory national standard, compliance is now required — effectively setting a unified access floor for L3/L4 in China and placing safety responsibility on automakers across the full vehicle lifecycle. For the industry, this amounts to roughly an 11-month compliance countdown, since vehicle validation cycles for clauses covering failure response and minimum-risk strategies are typically measured in years. Source: Securities Times, Sina Finance
US Teamsters Sue California DMV to Overturn New Driverless Heavy Truck Rules · autonomy
California Teamsters filed suit against the California DMV in Alameda County Superior Court on August 5, with the core allegation being that the agency permitted autonomous freight operations without completing the legally required "Standardized Regulatory Impact Assessment" (SRIA) — the DMV lifted its ban on autonomous driving for vehicles over 10,001 pounds gross weight on April 28. The union says the DMV underestimated the employment impact on California's more than 200,000 commercial truck drivers, and questions whether heavy driverless trucks have undergone sufficient public-road testing, seeking to have the rule invalidated and enforcement blocked. This is the first institutional obstacle driverless trucking has hit in its US commercialization push, and it takes a procedural-legality route rather than a safety-evidence one. Source: Transport Topics, The Hill
South Korea Eases Robot-Fencing and "One Site, One License" Restrictions, Unlocking KRW 4.2 Trillion in Investment · industrial
South Korea's Minister of Economy and Finance Koo Yun-cheol announced a package of deregulation measures the government expects to unlock KRW 4.2 trillion (about $2.96 billion) in investment, focused on freeing up already-ready projects stuck in the approval process, serving semiconductors, physical AI, and AI data centers as the three priority areas. The two measures directly tied to robotics are removing the "one site, one license" restriction and lifting the industrial robot fencing requirement — the latter determines whether collaborative and mobile robots can share a workspace with people, a direct constraint on production-line deployment density. Source: The Korea Herald, Asia Economy
Goldman Sachs: AI Investment Theme Shifting to Humanoids, 76,000 Units Projected for 2027, Scale-Up Not Until 2027–2029 · humanoid
Goldman Sachs' latest research report judges that as compute-infrastructure investment continues to expand, the AI investment theme is shifting from the compute race toward end applications, with humanoid robots as the next-stage growth engine; it projects global humanoid shipments reaching 76,000 units in 2027 and 502,000 units by 2032, with large-scale deployment only gradually unfolding from 2027–2029. The report also notes current applications remain largely proof-of-concept (POC), with scenarios extending into industrial handling, logistics workflows, and structured commercial settings. This timeline judgment diverges notably from figures reported in China: China's Ministry of Industry and Information Technology's figure this month of roughly 23,000 units shipped in the first half had over 70% still in showrooms and labs — the two figures actually point to the same underlying reality, that the shipment curve has taken off while the deployed-workstation curve has not. Source: Goldman Sachs research report, via Sina Finance
Counterpoint: Commercial Cleaning Robot Revenue Exceeds $400 Million in 2025, Pudu Robotics Leads in Both Shipments and Revenue · industrial
A Counterpoint Research report shows global commercial cleaning robot shipments exceeded 50,000 units and revenue exceeded $400 million in 2025, with Pudu Robotics ranking first globally in both shipments and revenue. The report also gives two directional figures: global commercial cleaning robot penetration remains below 10%, while annual shipments could exceed 400,000 units by 2030 at a compound growth rate above 50%. While humanoids are still debating their first stable deployment slot, cleaning robots — a single-task, structured-scenario category — have already built a countable revenue base, offering a ready reference point for observing when robotics starts becoming a real business.
Source: Counterpoint Research
This Week's Supply Chain: Money Flows to Both Transmission and Tactile Sensing, Dexterous Hands Still Stuck on Standardization · hardware
Stringing together this week's scattered signals, funding is concentrated almost entirely on two end-effector segments of the humanoid. On the transmission side: UOTECH's over RMB 200 million Series B with a 1.5 million-unit joint order backlog, NOUSBOT's near-RMB 100 million Series A+ expanding miniature planetary roller screw production with a stated 10-million-unit target, and Zhenghe Industrial's roughly RMB 700 million private placement redirected toward dexterous-hand components — the scale of capacity planning has jumped from "tens of thousands of units" to "millions of units," spanning harmonic joints to lead screws. On the tactile side, Sino Silicon Century placed an order this week for 200,000 units of tactile sensors with Flysource New Materials, sufficient by its own count to equip 40,000 industrial-grade dexterous hands — flexible tactile sensing is moving from samples to bulk orders. The sharpest point of tension isn't capacity but standards: industry statistics show dexterous-hand-sector funding in the first half already exceeded RMB 25 billion, more than all of 2025, yet interfaces, degree-of-freedom configurations, and lifespan specs vary widely across manufacturers — NOUSBOT's founder noted that most dexterous hands on the market in 2025 failed frequently after about 200 hours, and by 2026 there's still a gap to industrial-grade reliability. Component makers are building capacity for a whole-machine form factor that hasn't yet converged — the most real risk exposure in this week's supply chain. ⚠️ Compiled figuresSource: 36Kr, Eastmoney, 10jqka
Top comments (0)