Today's Highlights
· Waymo notified Uber it will launch its own app in Austin and Atlanta in January 2028, ending exclusivity in both markets
· Hyundai Motor announced its transformation into a "Physical AI Solutions Company," co-building a robotics reference platform with Nvidia
· Dongfeng Aeolus (Chinese automaker) × Hello Inc. (Chinese mobility platform) L4 robotaxis roll off the production line, with plans to deploy 10,000 units and expand to no fewer than 10 cities within two years
· Dexterous hands lifting 5 kg break after only about 50 cycles; even top products last just 300,000-500,000 cycles
· AGIBOT's (Chinese humanoid robotics startup) Hong Kong IPO could land as early as August, with cornerstone investors targeting a valuation of HK$40-50 billion
I. Research Progress
Nankai University's "one chip, three functions" ferroelectric transistor: squeezing sensing, computation, and activation into a single device · perception
Embodied vision systems have long been constrained by sensors, memory, computation, and activation circuits each operating in isolation — data must repeatedly shuttle between modules and undergo multiple analog-to-digital conversions, keeping both energy consumption and latency stubbornly high. A team led by Chen Xudong, Tian Jianguo, and Chen Shuqi at Nankai University, in collaboration with Kang Lixing's team at the Suzhou Institute of Nano-Tech and Nano-Bionics (Chinese Academy of Sciences), used a dual split-control-gate structure to modulate local polarization in CuInP2S₆, programming p-n/n-p/p-p/n-n homojunctions into an ambipolar WSe₂ channel. This lets a single ferroelectric transistor switch on demand between three modes — in-sensor computing, in-memory computing, and nonlinear activation — with 33 optical weight states, 32 multi-level conductance states, and ReLU/Sigmoid-type response times of about 8 ns. The team built end-to-end neuromorphic vision hardware on this basis, converting classification results directly into left-turn/right-turn/stop commands for a mobile robot, achieving 95% MNIST recognition accuracy, roughly 98% system-level simulation accuracy on traffic video processing, and 0.99 structural similarity between the full-hardware autoencoder reconstruction and the software benchmark. The work was published in Science Advances.
Li Yuan, Zhang Zhicheng et al. (Nankai University · Suzhou Institute of Nano-Tech and Nano-Bionics, CAS) · Science Advances 10.1126/sciadv.aed9436 · via: Nankai University Institute for Interdisciplinary Sciences source (WeChat, CN)
Jetson-PI: turning VLA from "demo-capable" into "continuously operable" on onboard compute · vla
Running π0.5 directly on Jetson Orin yields an inference time of about 1.4 seconds and a control frequency of only around 0.7 Hz, leaving the robot sluggish with visible pauses between action chunks — this is the most concrete wall facing edge-device VLA deployment. Peking University, AIRS, and PrimeBot Research Institute propose "look-ahead aligned asynchronous correction": rather than fusing old and new trajectories, they train a lightweight module of about 40M parameters (roughly 1% of the full VLA) to directly predict the future VLM environment representation after the current action completes, letting the action expert generate actions starting from that future moment — sidestepping the perception-execution misalignment inherent to asynchronous inference. The system also schedules by confidence: it skips the VLM and runs only the action expert during steady states, and only "looks again" when confidence drops at critical steps like grasping or placement. Without sacrificing accuracy, this raises the control frequency on Jetson Orin to 6.06 Hz, 8.66 times faster than native PyTorch inference.
Peking University · AIRS · PrimeBot Research Institute · arXiv 2607.12659 source · via: Sina Mobile source
Open Source · Tools · Benchmarks
· Jetson-PI: the Peking University team has open-sourced both the asynchronous inference code and the edge inference engine (PKU-SEC-Lab), targeting low-power onboard devices like the Jetson Orin/Thor source
· DexLink / SyLink Hand: material costs disclosed in the papers have dropped to the $400 range — DexLink drives 20 joints with 16 actuators in a 320-gram hand, completing all 33 Feix grasp types; SyLink goes further, driving 19 joints with just 11 actuators. The reference point is the LEAP Hand's sub-$2,000 cost in 2023, while the LEAP Hand V2's metal-palm durability variant actually rose to about $4,000 — the cost conversation is shifting from "how much does one hand cost" to tolerances, calibration, and lifespan source (WeChat, CN)
· ORCA Hand: material cost under 2,000 Swiss francs, with a publicly verified continuous run of 7 hours 17 minutes completing about 2,000 grasp-and-release cycles without tendon loosening or breakage — pushing benchmarking from "can it do it once" to "can it repeat it unattended" source (WeChat, CN)
II. Funding and Deals
AGIBOT (Chinese humanoid robotics startup) | Hong Kong IPO | Cornerstone target valuation HK$40-50 billion · humanoid ⚠️ Reported terms
Following the formal launch of its Hong Kong listing process on July 24, AGIBOT could complete its Hong Kong IPO as early as this August; cornerstone investors have set a target valuation of HK$40-50 billion (roughly RMB 34.6-43.3 billion), while AGIBOT itself had hoped to push the target valuation to around HK$80 billion. For reference, the company's most recently disclosed post-money valuation as of July 2026 was RMB 15 billion, with 2025 revenue exceeding RMB 1.05 billion — nearly a 20-fold increase year over year. According to Omdia, AGIBOT shipped over 5,100 units of general-purpose humanoid robots in 2025, roughly 39% of the global market, while rival Unitree Robotics (Chinese quadruped/humanoid robot maker) officially disclosed shipments exceeding 5,500 units over the same period — the discrepancy between the two figures illustrates that "world's number one" depends on how the statistics are scoped.Source: Zui Investment Bank source (WeChat, CN)
Jingling Zhikang (Chinese wearable prosthetics startup) | Angel round | Nearly RMB 100 million · embodied
Led by Longpan Investment, with participation from Jiangxi Financial Holdings, Xinhuadu Chairman Ni Guotao, and USTC-Silicon Valley (Hefei). This funding isn't targeting a dexterous hand mounted on a robot's end-effector — it's targeting one worn on a human hand. Hefei-based Jingling Zhikang makes a wearable bionic hand weighing about 370 grams that reads the wearer's movement intent to drive grasping motions. Price is the real battle it's fighting: imported high-end smart bionic hands typically cost RMB 300,000-500,000, and comparable Chinese products generally exceed RMB 100,000. The company — which completed a strategic acquisition of Lingxin Qiaoshou this past May — proposes first cutting the price to RMB 30,000-50,000, aiming to reach the RMB 10,000 price point within three years, with planned production of 500 units in 2026, 2,000 in 2027, and a push toward 10,000 in 2028. The company states its latest tests show a lifespan exceeding one million cycles⚠️ vendor claim.Source: Touyingli source (WeChat, CN)
Acrab (Singapore) | Cumulative funding | $350 million · hardware
Led by a $50 million investment from Huangpu River Capital (Chinese venture capital firm), with Temasek's Vertex making a heavy bet through Vertex Ventures SEA & India and Vertex Growth. Founder Ken Phua previously served as Arm's VP of Global Technology Strategy and Acquisitions and co-CEO of Arm China; the company is betting that "the CPU is becoming the center of the AI era again." Its first chip, the 5nm GΞLIX 1, has been announced — a heterogeneous architecture integrating a 20-core Arm CPU, a 3 TFLOPS GPU, and a dedicated NPU, delivering 650 TOPS of total AI compute. The company claims that in a Gemma 26B A4B configuration, prefill throughput reaches 1,416.8 tokens/s — 7.5 times that of an M4 Pro Mac Mini⚠️ vendor claim. Lead investor Huangpu River Capital has previously backed Montage Technology, Horizon Robotics, Galbot (Chinese humanoid robotics startup), and Paxini (Chinese tactile sensing startup).Source: LeiDi · Sina Finance source
Yisuo Intelligence | Seed round | Tens of millions of RMB · world-model
Exclusively invested by Zhengxuan Investment; the company simultaneously disclosed a term sheet for nearly RMB 100 million. As world models collectively pivot toward robotics and autonomous driving, Yisuo Intelligence is targeting scientific world models instead: its Iso-SWM engine + Iso-Sciland research workspace + Iso-LabOS experimental feedback system run a "predict-experiment-calibrate-repredict" wet-dry closed loop, starting in life sciences before expanding into materials, energy, chemical processes, and semiconductors. The reference point is Isomorphic Labs (Alphabet subsidiary)'s $2.1 billion funding round completed this May — this segment has clearer customers and willingness to pay, but requires a combination of AI algorithms, domain knowledge, and wet-lab engineering capability.Source: Rongleme source (WeChat, CN)
III. Commercialization and Deployment
Dongfeng Aeolus × Hello Inc.: L4 robotaxi rolls off production line, plans to deploy 10,000 units · autonomy
Following the delivery of Hello Inc.'s first mass-produced robotaxi, the HR1, yesterday, the vehicle officially rolled off the line at the Dongfeng Nissan Huadu plant on July 24 — the project was jointly built by Dongfeng Aeolus and Guangzhou Zaofu Intelligence (Hello Inc.'s inclusive mobility arm), based on the Aeolus VX6 electric SUV, with plans to deploy 10,000 units at scale. The first batch will run routine demonstration operations in Guangzhou and Shenzhen, with plans to expand to no fewer than 10 cities nationwide between 2026 and 2027. What's genuinely worth noting is the model itself: "automaker mass production + mobility platform's in-house autonomous driving" — shifting robotaxis from the retrofit-vehicle route to a product coming straight off an automaker's production line. The VX6's cabin volume can accommodate lidar and onboard computing units, and its economics have already been validated through high-intensity ride-hailing operations.Source: Caijing 21 source
FAW Jiefang delivers 73 L2 navigation-assisted heavy trucks to Annto Logistics, targeting "two drivers to one" · autonomy
On July 24 in Changshan, Zhejiang, FAW Jiefang delivered the first batch of 133 new trucks to Annto Logistics, of which 73 units are J6E L2 navigation-assisted heavy trucks entering operation on the national trunk-line transport network. The company states the model, equipped with its self-developed "vehicle + assisted driving" system, can achieve fuel savings of 3% to 5% and enable a shift from "two drivers to one" on mid-to-short-haul trunk routes of 500-1,000 kilometers — this is a direct cost line item for trunk-line logistics, not autonomous driving. As of June 2026, Annto operated over 11,000 company-owned trucks and trailers, with network coverage reaching about 99.7% of counties nationwide, and 1,376 FAW Jiefang trucks now in operation on its network.Source: Shanghai Securities News · Sina Finance source
South Korean Navy uses a humanoid robot as a ship's helmsman for the first time, Pibot passes phase-one testing · humanoid ⚠️ Simulator stage
The South Korean Navy, in partnership with KAIST (Professor Shim Hyun-chul's team), placed the humanoid robot Pibot in a helmsman role for the first time in a combat exercise: when the officer of the watch ordered "left rudder 5 degrees, course 330," the robot repeated the command, turned the wheel, and — once the simulated ship stabilized — reported "steady on course 330," completing the full repeat-back, steer, and report cycle a human helmsman would follow, with command comprehension driven by a large language model. The test spans four phases, with this being the first-phase land-based ship-handling simulator; subsequent phases will cover in-port mooring, daytime sea navigation, and continuous day-night navigation, with researchers analyzing time-to-completion and steering precision from command to execution. The South Korean Navy's motivation is stated plainly: responding to the shrinking pool of military recruits from population decline.Source: UPI source
Weave Robotics prices a mobile humanoid robot at $7,999 for the consumer market · humanoid ⚠️ Preorder terms
Its first mobile humanoid robot, Isaac 1, comes with a full price of $7,999 and a $449/month payment plan, with first shipments set for fall 2026. The company has already shipped its home laundry-folding robot, Isaac 0, within California, and sold several Isaac robots to customers in the San Francisco area. The price genuinely pulls humanoid robots into the psychological range of premium home appliances — but it's worth noting that what's shipped so far is the task-scoped Isaac 0, not the mobile humanoid Isaac 1. The two aren't in the same tier of mechanical complexity or home-safety requirements; what's being sold right now is trust in the first delivery batch, not off-the-shelf capability.Source: Zhenji Zhineng source (WeChat, CN)
Shanghai Electric unveils embodied factory robots and an "AI-native" factory concept at WAIC · industrial ⚠️ Demo terms
This industrial conglomerate, known for heavy equipment and power infrastructure, unveiled a full lineup of embodied-intelligence robots for production lines at WAIC 2026, along with an "AI-native" smart factory model — not layering AI onto legacy equipment, but having planning tools, robots, machine tools, and control rooms share the same data stream from day one. The positioning is that of an integrator capable of delivering robots, software platforms, and monitoring systems as a bundle. What's conspicuously absent from the public disclosure is exactly what buyers would ask first: pricing, a commercial rollout timeline, and the number of pilots currently running.Source: Mobilités Urbaines source
IV. Industry Developments
Waymo notifies Uber: self-owned app launching in Austin and Atlanta in January 2028, ending exclusive partnership · autonomy
An Uber spokesperson confirmed to CNBC that the company has received notice from Waymo that it will launch its own app in Austin and Atlanta in January 2028, running alongside the existing Uber channel. The two companies' contracts covering these markets run at least through May 2028, during which hundreds of Waymo robotaxis will remain on the Uber platform — and Uber, in turn, gains room to bring in other autonomous vehicle fleets in both cities. This isn't an isolated move: the two companies already ended their partnership in Phoenix earlier this year, Waymo now operates independently in 9 other markets, and it ran a non-exclusive partnership with Lyft in Nashville last year. The Financial Times reports that Waymo has internally discussed a complete split from Uber, with friction points including conflicting regulatory proposals the two companies have pushed in different states; Uber's CTO published a video earlier this year publicly criticizing Waymo vehicle behavior as "dangerous and unsettling," and in May, Uber's CEO questioned — without naming names — how autonomous vehicles perform around school zones and in emergencies on an earnings call. Uber's stock fell more than 4% the day the news broke (per CNBC; other financial outlets reported a drop of about 9% that day — the figures diverge). The fight over robotaxi distribution rights is shifting from "whose vehicles are safer" to "who controls the passenger gateway."Source: CNBC source · TechCrunch source
Hyundai Motor Group announces transformation into a "Physical AI Solutions Company," co-building a robotics reference platform with Nvidia · autonomy
Hyundai Motor Group Executive Chair Chung Euisun laid out the group's Physical AI roadmap at an AI summit in San Francisco, attended by around 150 South Korean corporate executives, U.S. tech company executives, entrepreneurs, and students. In his words, the group is "moving beyond the boundaries of traditional automaking, expanding into autonomous driving, robotics, and AI-defined factories, and accelerating its transformation into a Physical AI Solutions Company." The vision progresses across three layers: smart devices such as vehicles and robots → smart spaces such as AI factories → and ultimately city-level integrated intelligence. Hyundai's stated point of differentiation is a data flywheel — using large-scale manufacturing, mobility, robotics, and service operations to continuously feed real-world industrial operating data back into its AI. On the partnership front, beyond co-developing a "Robot Reference Platform" with NVIDIA, the group also cited its partnership with Waymo and Boston Dynamics' strategic collaboration with Google DeepMind, and said it will invest in local ecosystem projects such as the Saemangeum AI Valley.Source: PR Newswire source
Dexterous hands remain the real bottleneck for mass production: hands lifting 5 kg break after about 50 cycles, top products last only 300,000-500,000 cycles · hardware
A robotics company executive who has already shipped to partners (using the pseudonym Wang Hao) told Yicai that the Yinshi Robotics dexterous hand his company procured broke after just about 50 cycles of handling objects around 5 kg in a warehousing and logistics scenario; given that these hands cost under RMB 10,000, he considers this "understandable" — but the bar for industrial use is tens of thousands of cycles without failure, since once a robot is deployed in a factory, hands can't be swapped out every time one fails. Yinshi responded that its load-bearing tests have reached up to one million cycles, that Wang Hao's case was an isolated incident "not particularly representative," and disclosed that the company shipped over 10,000 units of dexterous hands last year to around 700 customers⚠️ conflicting claims. A harder data point comes from Geek+ (Chinese warehouse robotics company): Dong Qinpeng, head of its embodied intelligence R&D, says warehouse robots need to grasp 8,000 to 10,000 times per day, while today's leading dexterous hands on the market last only 300,000 to 500,000 cycles — and that's still a low-degree-of-freedom product — meaning "it might break down after about a month of use." A compromise solution has already emerged: Geek+'s recently launched robot switched to a three-finger dexterous hand, covering the massive SKU variety of warehousing while leaving room for reliability and lifespan design. The gap between the elegant five-finger hands on display floors and the practical two-finger grippers on production lines is precisely the cost of a supply chain that hasn't yet been battle-tested.Source: Yicai source (WeChat, CN)
Nvidia draws Japanese industrial giants including Toyota and FANUC into the physical AI camp · hardware
Jensen Huang met this week with leaders of Japanese industrial companies including Toyota, Fujitsu, Kawasaki Heavy Industries, FANUC, and Kioxia to discuss deploying physical AI in their factories, saying "the next frontier of AI is the physical world, and this is a once-in-a-century opportunity for Japan." Kawasaki, FANUC, and Yaskawa — three robotics and automation manufacturers — already use Nvidia technology. The scale comparison is worth noting: under Nvidia's new segment reporting, data center revenue reached $75 billion last quarter, up 92% year over year, while the newly created edge computing segment (robotics, automotive, PCs) brought in only $6.4 billion, up 29%. Management expects Blackwell/Rubin chips to generate $1 trillion in revenue between 2025 and 2027 — physical AI can't yet carry the stock price on its own in the near term, but it's the next stop in this full-stack (DGX + Jetson + Cosmos) strategy.Source: The Motley Fool · Yahoo Finance source
New York school district halts "robot teacher" program after state education department pressure · adjacent
The Salamanca school district in western New York State announced Friday it is pausing its pilot bringing a humanoid robot from Realbotix into classrooms, citing the need to finalize student data privacy agreements with the state education department and to communicate further with the community. The pause came hours after State Education Commissioner Betty Rosa sent a letter stating "the department remains concerned about the potential impact of humanoid robots within high school classrooms." The project cost $57,590, was approved by the local school board last month, and had planned to introduce the robot — named Sally — this fall: it features silicone skin, can perform upper-body movements and facial expressions while remaining seated, and comes paired with an AI teaching-assistant avatar students can access on their laptops. Realbotix's CEO said the company respects the school's decision. Public acceptance of humanoid form factors in public-service settings is emerging as a constraint independent of technological maturity.Source: New York Focus source · The Guardian source
BYD officially enters humanoid robotics, with in-store sales assistance as its first deployment scenario · humanoid ⚠️ Statement only
BYD has confirmed it is developing humanoid robots in-house, with the first deployment scenario being in-store smart sales assistance — handling standardized services like product explanations and customer reception. This echoes an earlier statement from the company's Executive Vice President Li Ke, who said the robots could eventually be sold to consumers at 4S dealerships once the technology matures. There is currently no verifiable disclosure of the product's form factor, mass-production timeline, or deployment scale; a previously circulated figure on specific deployment numbers was denied by the company as inaccurate.Source: Autohome source
Hikvision Robotics reports H1 revenue of RMB 4.028 billion · industrial
The machine vision and mobile robotics business reported first-half revenue of RMB 4.028 billion. Amid the dense flow of humanoid-robotics narratives, the actual revenue scale of industrial machine-vision/AMR companies like this one serves as a yardstick for how far "embodied intelligence" still is from sustainable cash flow.Source: EEFocus source
Hardware · Supply Chain
· Dexterous hand demand: per GGII data, dexterous hand sales in China are projected to surge from 19,200 units in 2025 to 70,200 units in 2026, an increase of over 260%; the resulting downstream volume is pushing capital into segments like micro coreless-cup motors and ultra-thin reducers previously dominated by European, U.S., and Japanese manufacturers source (WeChat, CN)
· Planetary roller screws: research reports state that a single humanoid robot requires 14 planetary roller screws (for linear joints) and 12 miniature ball screws (for dexterous hands), with planetary roller screws accounting for 19% of value — the single highest-value core component in the whole machine⚠️ research-report claim source (WeChat, CN)
· Aoyi Technology: dexterous hand orders surpassed 10,000 units in H1 (previously reported); newly disclosed that its next-generation product will launch in August, still using a linkage design with 16 degrees of freedom, striking a balance between heat dissipation and transmission efficiency source (WeChat, CN)
· Axera: announced the formation of a wholly owned subsidiary, "Axera Compute," extending further into the compute side source
Top comments (0)