Today's Highlights
· Unitree's Shanghai STAR Market IPO is imminent (listing on August 19), with a new humanoid robot "Superman" launched the same day
· Force sensor maker Bluepoint Touch Control raises several hundred million RMB in Series D, holding a 72.6% share of the Chinese humanoid market
· Gravis Robotics raises $200 million from SoftBank, the largest-ever Series A in construction robotics
· China's National Bureau of Statistics: industrial robot output up 28.5% year-on-year over the first 7 months
· Hunan province sets a target of over RMB 200 billion in intelligent robotics industry-chain revenue by 2030
· Serve Robotics integrates with Grubhub, simultaneously rolling out micro-depots and Moxi 2.0
· Zhixiang Future's (Chinese AI startup) interactive world model tops the WBench navigation track with a score of 80.9
1. Research Progress
Turning VLA into a tool-calling agent: 30,000 trajectories buy a 20% jump in success rate · vla
End-to-end VLA models suffer from a continuous, enormous action solution space, forcing generalization to be squeezed out through sheer data volume. ART takes a different approach: rather than modifying the action head, it injects ready-made tool modules — low-level vision, high-level affordance, and embodiment augmentation — into any VLA, letting the model call them on the fly at inference time and slicing the continuous solution space into composable discrete calls. The training set uses only 30K tool-call and demonstration trajectories, far smaller than comparable baselines, yet still outperforms mainstream baselines by 20% in success rate across simulation and real-robot tasks, including pick-and-place under low light from novel viewpoints.
Yi Ding et al. · arXiv 2608.14047 source
ReflexBench: existing VLA benchmarks barely test "reaction speed" · vla
Mainstream manipulation benchmarks mostly test static placement tasks, leaving dynamic interaction scenarios largely untested — yet in real workstations, objects slide, fall, or get pushed away. ReflexBench fills this gap with 6 dynamic tasks and decouples simulation stepping from robot control, allowing configurable latency under synchronous/asynchronous inference. The accompanying ReflexVLA skips large-scale real-robot pretraining, instead improving temporal reasoning via implicit future prediction within the visual backbone and multi-frame temporal fusion, while compressing deployment latency through batched visual encoding and CUDA Graph replay — static-task accuracy holds steady while dynamic performance improves.
Yuxuan Chen et al. · arXiv 2608.14379 source
AdvDex uses adversarial learning to strip "embodiment appearance," transferring human-hand demonstrations zero-shot to robot hands · manipulation
Scaling dexterous manipulation is bottlenecked by the high cost of real-robot demonstrations and the mutual incompatibility of different embodiments' action spaces; policies trained on mixed data also entangle task cues with embodiment appearance. AdvDex offers three components: a large-scale human-hand demonstration dataset called OmniShare (with kinematic supervision and tactile measurements); a unified JAAS action representation that maps human hands, dexterous hands, and parallel grippers into a common SE(3) wrist pose + 15 finger joints format; and domain-adversarial learning to strip embodiment-specific features from visual representations. The result is usable zero-shot human-to-robot skill transfer, with few-shot adaptation to unseen objects and environments.
Zhiyue Zhao et al. · arXiv 2608.14028 source
hint²: using a hierarchical world model at inference time to bend short-sighted policies toward long-horizon logical constraints · world-model
Language-conditioned policies generate short-horizon action chunks with closed-loop replanning, while linear temporal logic (LTL, used to express non-Markovian instructions like "do A before B, and never C") can only be evaluated over long-horizon trajectories — an inherent mismatch. hint² has a high-level world model predict how actions will advance the atomic propositions of an LTL automaton, while a low-level dynamics model handles immediate safety, with the two levels dividing guidance responsibilities. It outperforms existing inference-time guidance methods on CALVIN and completes instructions with complex liveness and safety constraints on a real UR5e robot.
Moritz Zoellner et al. · arXiv 2608.13678 source
PRM-as-a-Judge 1.5: stop evaluating embodied models with a single "success rate" number · benchmark · HF↑8
Binary success rate obscures too much information — how far a task progressed before failing, whether the system can recover after a fall, and the execution quality of a successful run all get compressed into a 0 or 1. This toolkit converts rollout videos into dense progress curves, and on top of the 1.0 version adds three new metrics separately characterizing failure-side progress, recovery after setbacks, and success-side execution quality, while using RoboPulse++ to in turn check whether the process reward model itself is trustworthy. The authors also release a complete evaluation suite including the benchmark, metric implementations, and visualization tools.
Yuyang Liu et al. · arXiv 2608.14284 source
FlatLab: a unified framework and simulation benchmark specifically for "flat objects that can't be grasped" · manipulation
The difficulty with paper, cloth, thin boards, and similar objects is that the initial pose offers nowhere to grip; existing approaches mostly rely on heuristic pre-manipulation and are evaluated only in closed settings. This work splits the task into a policy generator and an action-execution module: the former learns object-agnostic, policy-centric representations from point clouds using simulated data transformations and contrastive learning; the latter decomposes long-horizon manipulation into reusable action primitives that are dynamically composed. The accompanying FlatLab benchmark provides high-fidelity physical simulation and standardized task definitions for both rigid and deformable flat objects, with code publicly released.
Xingyu Zhu et al. · arXiv 2608.14049 source
Using cheap simulated "surrogate systems" to predict real-robot failures, unearthing twice as many failure cases at the same budget · benchmark
Failures in autonomous systems are rare and heterogeneous, while real-robot testing budgets are limited; simulators or low-fidelity surrogate systems can sample at massive scale, but the sim2real gap often means surrogate failures don't transfer. This method applies control-variate-style residual correction to surrogate failure signals, learning a local predictor of target-system risk, paired with a support-aware mutual-information objective that favors both realistic, trustworthy regions and broad coverage of failure modes. Across three task categories — autonomous driving, manipulation, and quadruped velocity tracking — it discovers up to 2x as many failures as random sampling and active-learning baselines, covering severe failure modes that competing methods miss.
Anjali Parashar et al. · arXiv 2608.13719 source
SSP: only by running the same conflict event across synthetic, simulated, and real-vehicle domains can you tell if the gap is real · autonomy
Cross-domain evaluation of autonomous-driving VLA models commonly uses separately chosen datasets for each domain, making it impossible to tell whether measured gaps reflect domain sensitivity or simply different scene content. SSP starts from synthetic long-tail videos and fixes an event specification — road topology, participant roles, relative motion, conflict evolution, passage order, response constraints, and event phase — then reproduces the same event in both CARLA and a closed test track, passing a transfer audit before evaluation. Across two scenario types, cut-ins and vulnerable-road-user crossings, the composite capability scores across the synthetic/simulated/physical domains are 0.259 / 0.291 / 0.325 respectively — but the best-performing domain varies by scenario, undermining the default assumption that "real vehicles are always harder and more accurate."
Haojie Feng et al. · arXiv 2608.14024 source
Other papers today: Graph-MambaNav feeds object-relation knowledge into a spatio-temporal graph Mamba for object-goal navigation (arXiv 2608.13723 source); OccPlanner conditions a diffusion planner on occupancy grids to solve pixel-goal navigation (arXiv 2608.14160 source); OpenBelief-Nav preserves minority hypotheses rather than prematurely committing to semantic labels in open-vocabulary scene graphs (arXiv 2608.13923 source); GhostPoint performs self-supervised pretraining by "hallucinating" occluded LiDAR structure (arXiv 2608.14428 source); BICPO-VLA performs continuity preference optimization for the request-handoff gap in asynchronous VLA (arXiv 2608.13924 source); Rollplex enables shared cross-phase GPU use across rollout, scoring, and training stages of VLM post-training (arXiv 2608.14498 source); Onto-EV-WM adds ontology-grounded failure diagnosis and verification-gated repair to world models (arXiv 2608.13901 source); the effect of yarn-twist structure on pressure and proximity sensing properties of fabric capacitive sensors for robot skin (arXiv 2608.14406 source); MMUSV-Sim, a simulation and data-generation platform for multi-USV cooperative perception (arXiv 2608.14207 source); information-theoretic path planning for Mars exploration based on expected free energy (arXiv 2608.14466 source).
2. Funding & Deals
Bluepoint Touch Control (Chinese force-sensor maker) ｜ Series D ｜ Several hundred million RMB · hardware
Led by GAC Capital, the investment arm of GAC Group, with participation from Deye Co. and Xichen Capital. Founded in 2019, the company's core team comes from the China Aerospace Science and Technology Corporation, the Chinese Academy of Sciences, Beihang University, and Northwestern Polytechnical University, transferring high-precision force control from aerospace applications to robotics. According to the "2025 China Humanoid Robot Six-Axis Force Sensor Market Analysis Report" from the GGII research institute, its six-axis force sensors hold a 72.6% share of the Chinese humanoid robot market, with mass adoption by Zhiyuan Robotics (Chinese humanoid startup), Galbot (Chinese robotics startup), Xiaomi Robotics, UBTECH, Engine AI, and AgiBot (Chinese embodied AI startup). Following this round, its shareholder base spans battery makers (CATL), automakers (GAC, SAIC), robot manufacturers (Zhiyuan Robotics, Galbot), and semiconductor manufacturing (SMIC) — Vice President Zhang Pengling said Zhiyuan Robotics and Galbot were already core customers before investing, going "from customer to shareholder." On the production side, the company has built China's first fully automated force-sensor production line in Guangdong, with a designed annual capacity of 1 million joint force sensors and 200,000 end-effector six-axis units, cutting delivery time from overseas brands' 6–8 weeks to 2–3 weeks.Source: 36Kr Hard Tech source
Gravis Robotics (Switzerland) ｜ Series A ｜ $200 million · industrial
Following reports of SoftBank's acquisition talks in July, this round was funded solely by SoftBank Group, with the company calling it the largest-ever Series A in construction robotics. Founded in 2022 as a spinoff from ETH Zurich, the company retrofits existing heavy excavators with autonomous driving capability — its Gravis Rack control system can be installed after the fact on machines from Caterpillar, John Deere, JCB, and Hitachi. The technical challenge runs opposite to typical physical AI scenarios: autonomous driving and robotic arms aim to avoid disturbing the environment, whereas an excavator's job is literally to disturb terrain. Co-founder and CTO Dominic Jud said the model learns in simulation from the engine sound, vibration, and hydraulic resistance that human operators rely on to "read the soil," translating this into machine telemetry signals that respond to underground soil-force changes at microsecond speed.Source: SiliconANGLE source
MiFeng Technology ｜ New round ｜ Several hundred million RMB · adjacent
Led by China Telecom, with participation from Zhangjiang Group, and existing shareholders Sequoia China and Yuanqi Innovation adding beyond their pro-rata share — this is the third funding round in half a year for this physical AI data platform, founded only in early 2026. The funds go toward scaling production of its MEgo series of body-free data-collection devices and toward tens-of-millions-of-hours-scale physical interaction data capacity. The company frames its industry benchmark as: mainstream large language model pretraining data has already reached 100 trillion tokens, equivalent to roughly 10 billion hours, while physical AI data has lower information density and higher noise, meaning demand will only be greater. Its MEgo Engine governance platform claims to boost traditional manual data-processing efficiency by over 10x, and has already supplied training data for embodied foundation models such as Ant Group's LingBot-VLA 2.0.⚠️ Vendor claimSource: DoNews source
Yuansheng Technology (Shenzhen) ｜ Pre-A / Pre-A+ / Pre-A++, three rounds ｜ Over RMB 100 million cumulative · hardware
Founded just over a year ago, this embodied-tactile-sensing company completed three rounds led respectively by Hengxin Huaye, Shangfu Electric, and Topband (601689), with participation from Xichen Capital and others — two of the three lead investors are automotive sensor and components players. The company disclosed that its first multi-dimensional tactile standard product has entered projects with several dexterous-hand, robotics, and data-collection customers in China and abroad, with total sensor orders this year reaching the hundreds of thousands of units; a joint-venture mass-production line for multi-dimensional tactile sensors built with Shangfu Electric went into operation on August 10. The tactile-sensing sector has seen a wave of funding recently (previously reported); the new information here is the production line going live and the order volume — i.e., moving from sample validation to mass-production sourcing.Source: Sohu source
Digital Huaxia (Shenzhen) ｜ Strategic round ｜ Hundreds of millions of RMB · humanoid
Investment from multiple industrial capital investors, with existing shareholder Cowin Capital adding to its stake. Founded in March 2024, the company focuses on bio-inspired humanoids and scenario intelligence; its new-generation bio-inspired humanoid Xialan R03, unveiled at WAIC, supports multilingual, wake-word-free conversation for reception and guided-tour use cases. Funding will focus on scaling production of its bio-inspired humanoid series and its scenario "brain" platform RoboEase — the latter replaces coding with graphical workflow orchestration and has already been deployed in banks, senior-care facilities, government service halls, hospital triage, and hotel front desks.Source: PEdaily source
Robotera (Xingdong Jiyuan) ｜ Planned Hong Kong IPO ｜ Up to $1 billion raise · humanoid ⚠️ Rumored
No filing has been made yet; reports put the fundraising size at up to $1 billion (roughly HK$7.8 billion). Founded in August 2023 as a spinoff from Tsinghua University's Institute for Interdisciplinary Information Sciences, the company is led by founder Chen Jianyu, an assistant professor at Tsinghua, and says it is the only humanoid robot company in which Tsinghua holds a stake; its flagship product is the full-size bipedal L7, alongside the M7 upper-body humanoid, the wheeled Q5, and the XHand dexterous-hand series. On deployment, the company says its robots are already in use across more than 10 logistics centers of China Post and SF Express. On funding timeline, Caixin reported in March that it completed a RMB 1 billion strategic round at a valuation exceeding RMB 10 billion; in April, a separate round of over $200 million led by SF Holding also closed, with participation from Sequoia, IDG, and CICC Capital. It joins Zhiyuan Robotics, Engine AI, and Rokae in choosing Hong Kong over a US listing.Source: Startup Fortune source
Takt (United States) ｜ Series A ｜ $9.25 million · industrial
Funds will go toward expanding its warehouse intelligence platform.Source: citybiz source
3. Commercialization & Deployment
Serve Robotics rolls out Grubhub and Wonder integrations alongside micro-depots · adjacent
Serve Robotics (SERV) filed an 8-K on August 17, simultaneously announcing a delivery partnership with Grubhub, deployment with Wonder, entry into new markets, rollout of micro-depots, and launches of a merchant-facing product called Beacon, an advertising product called Characters, and an upgrade to its hospital robot, Moxi 2.0. The economics bottleneck for sidewalk delivery robots has always been density — how many orders a single robot can complete in a day depends on how close it is to pickup points; micro-depots and new aggregator-platform integrations target exactly this, rather than any new locomotion capability.Source: TradingView / GlobeNewswire source
Pudu Robotics takes the global No. 1 spot in both shipment volume and revenue for commercial cleaning robots · industrial
A Counterpoint Research report released August 13 shows that global commercial cleaning robot shipments exceeded 50,000 units in 2025, with market revenue surpassing $400 million; Pudu Robotics ranked first globally in both shipment volume and revenue, the only company to top both metrics simultaneously. A separate Frost & Sullivan report put its global market revenue share at roughly 29%. On the customer side, it has already landed chains including Walmart, Edeka, and Denner, covering markets in Europe, the US, Japan, and South Korea. The research firm also assessed current global penetration at under 10%, projecting annual global shipments to exceed 400,000 units by 2030. The company's small scrubbing robot, the PUDU ET1, launched August 12, targets store spaces of 100–800 square meters.Source: kkj.cn source
Yikong Zhijia signs overseas cooperation MOU with XCMG; over 260 XCMG mining trucks already run its driverless system · autonomy
The agreement was signed at an event in Perth celebrating the company's Hong Kong listing and its Australian milestones, pointing toward markets including Australia. The existing installed base is substantial: within China, over 260 XCMG mining trucks are already running Yikong Zhijia's ORCASTRA driverless transport system, with over 100 more awaiting deployment; in Australia, the company has already automated 6 Komatsu HD1500 rigid-frame mining trucks, operating around the clock at the Havana pit of Norton Gold Fields. XCMG also delivered its first batch of 12 XDE260 large rigid mining trucks to Western Australia earlier this year, and has supply or framework agreements with Fortescue, BHP, and Rio Tinto respectively. Elaine Jin, Yikong Zhijia's Chief Growth Officer for Australia, said close collaboration with OEMs allows its autonomous-driving components to be installed directly on the assembly line, greatly simplifying deployment.⚠️ PlannedSource: International Mining source
A snapshot of Jiangsu's data-collection factories: Galbot's Suzhou base runs 100 robots training across 20 simulated scenarios · adjacent
In a 5,000-square-meter space at the Jiangsu International Data Port Embodied Intelligence Innovation Center in Nanjing's Xuanwu District, 50 data-collection workers operate daily in 1:1 replicas of supermarket, restaurant, and sorting-facility environments. At the Taihu Embodied Intelligence Industrial Park in Suzhou's high-tech zone, Galbot's center spans roughly 11,000 square meters, with 100 robots training simultaneously across 20 simulated scenarios; the company says over 99% of its data is synthetically generated, producing a billion-scale manipulation dataset per week at a cost of just one-hundredth of real-world collection. Meanwhile, in Wuzhong District, Magic Atom's (Chinese robotics startup) data-collection factory produces roughly 16,000 valid data points daily, with real-world data accounting for over 80%. JD.com is taking a different route in Suqian — residents wear a 220-gram JoyEgoCam device while doing housework, collecting data incidentally, with plans to mobilize over 100,000 internal employees and 500,000 external participants within two years to gather over 10 million hours of data. Cost remains the sector's real bottleneck: 30 seconds of real-robot manipulation data costs RMB 10–15, or up to a thousand yuan per hour, with industry estimates putting the cost of accumulating 200,000 hours at RMB 200 million. According to GGII, at least 15 data-collection and training factories have been built or are under construction across China.Source: Xinhua Daily · Jiaohuidian source
AiMOGA passes 2,000 cumulative overseas deliveries, reaching over 60 countries and regions · embodied
Chery-owned AiMOGA held a global delivery ceremony in Wuhu, announcing cumulative overseas deliveries exceeding 2,000 units, with products and services now in over 60 countries and regions; in Malaysia, the products will serve showrooms for Chery, OMODA&JAECOO, iCAUR, and Lepas. Chery International President and AiMOGA President Zhang Guibing emphasized that "going overseas isn't just about product capability — it's an end-to-end capability spanning R&D, manufacturing, supply chain, delivery, service, and localized operations." The company also disclosed that it deployed 110 smart traffic-police robots across multiple Chinese cities in the first half of the year. Showroom reception and traffic guidance remain relatively low-complexity service scenarios, but the 2,000-unit scale itself indicates this form factor has already achieved repeat purchases.Source: PR Newswire source
HarmonyOS Intelligent Mobility's Zunjie V800 and V680 receive L3 autonomous-driving test licenses in Hefei · autonomy
Two Zunjie models from HarmonyOS Intelligent Mobility have obtained L3 test licenses in Hefei. This is the first vehicle-level move reported publicly since the mandatory national L3/L4 standard was finalized earlier this week (previously reported); a test license is still a full approval process away from mass-production road use.Source: Sina Auto source
Imagry and UNVI launch high-definition-map-free autonomous bus in Germany · autonomy
The two companies launched an autonomous bus solution in the German market that does not rely on high-definition maps, instead using a pure-vision, real-time mapping approach — a contrast to the mainstream Chinese approach for autonomous buses, which relies on high-definition maps combined with vehicle-road cooperation.Source: CNFOL source
4. Industry Developments
Ahead of its IPO, Unitree launches "Superman," with a standing jump of 2 meters and top speed of 12.66 m/s · humanoid
Unitree Robotics launched a new humanoid robot called "Superman" on August 17, with official figures citing a leg length of 0.85 meters, a standing vertical jump of about 2 meters, and a top running speed of 12.66 m/s — the company says both figures exceed corresponding human limits. For reference, the human standing high jump world record is roughly 1.8 meters, and Usain Bolt's peak speed mid-100m is roughly 12.42 m/s. Founder Wang Xingxing said the entire unit went from project initiation to launch in just over 3 months and is still in rapid iteration, with substantial room left to improve motion control and stability. The product complements the earlier G1 and H1 lines, continuing Unitree's focus on high-dynamic locomotion. It's worth distinguishing: these are official peak motion figures, neither independently verified nor equivalent to payload-handling capability — the gap between peak explosive power and "being able to actually do work" involves entirely different engineering problems. The timing is also no coincidence: the company announced the same day that its stock will list on the Shanghai STAR Market on August 19 (continuing its earlier IPO process), with an issue price of RMB 150.80 per share and an issue-based market capitalization of roughly RMB 60.993 billion, with 27,666 shareholder accounts post-issue; China's A-share robotics sector saw a volume spike in the afternoon on the news.⚠️ Vendor claimSource: The Paper source, Robot Lecture Hall source
China's National Bureau of Statistics: industrial robot output up 28.5% year-on-year over the first 7 months · industrial
Fu Linghui, spokesperson for China's National Bureau of Statistics, said at an August 17 press briefing that equipment manufacturing and high-tech manufacturing showed strong growth momentum in January through July, with industrial robot output up 28.5% year-on-year, and 3D printing equipment and lithium-ion battery output up 52.3% and 40.2% year-on-year respectively. This is an output figure, not shipment or installation volume — read alongside the customs data released earlier this month showing 7-month industrial robot exports up 13.2% (previously reported), it suggests most of the growth is still being absorbed within China.Source: Xinhua News Agency source
Hunan issues robotics industry implementation guidelines: targeting over RMB 200 billion in industry-chain revenue by 2030 · industrial
Document No. 42 [2026] from the Hunan Provincial Government sets two-tier targets: by 2028, provincial intelligent robotics industry-chain revenue is targeted to exceed RMB 100 billion, with roughly 5 chain-leading enterprises cultivated; by 2030, revenue is targeted to exceed RMB 200 billion, with roughly 10 chain-leading enterprises, placing the province's overall industry strength in the national first tier. The document lists ten key initiatives, the most distinctly Hunan-specific of which is "converting existing manufacturing strengths" — guiding construction machinery, rail transit equipment, and aerospace enterprises to extend their precision machining, advanced die-casting, and final-assembly integration capabilities into robot structural parts, joint modules, and high-precision reducers. Regional division of labor is spelled out in detail: Changsha handles main bodies, high-end controllers, and embodied-intelligence algorithms; Zhuzhou/Xiangtan/Hengyang handle motors and electronic controls plus precision joints; Shaoyang handles wiring harnesses and joint motors; Yueyang and Yongzhou handle servo motors; Loudi handles precision cast and forged components. On the application side, the plan commits to supporting roughly 20 flagship use cases and roughly 30 new products annually, and requires state-owned enterprises to take the lead in opening up use cases. The real effectiveness of regional industrial policy ultimately depends on where fiscal support and orders actually land, but the shift in emphasis from "attracting whole-unit makers" to "supplementing components and materials" aligns with the recent trend of early-stage capital moving downstream into lead screws and force sensors.Source: Hunan Provincial People's Government General Office source
Zhixiang Future launches interactive world model HiDream-O1-World, tops WBench navigation track · world-model
Users can enter a generated scene and move freely, switching between first- and third-person perspectives, and can direct characters to grab, run, crouch, or drive using text, voice, or images, altering the course of the story. The key difference from video models lies in "what happens after generation": the model must remember prior spatial structure and character relationships and stay consistent across multiple rounds of interaction. Its approach separately processes a scene's geometric structure and appearance, layering on two mechanisms — 3D-prior-injected Memory context and Test-Time Training — to maintain long-term consistency; the foundation is a proprietary unified Transformer (UiT) architecture that maps pixels, text, video, audio, actions, and spatial relationships into a shared token space. On the WBench leaderboard's Navigation Split, released by Meituan's LongCat team and Fudan University, the model scored 80.9 overall, ranking first among 30 evaluated models, with a consistency score of 88.0 and a physics score of 73.3 (also first place in the physics dimension) — the leaderboard is a third-party evaluation, but still represents testing under bounded conditions, and how well the model performs in open-ended scenes over longer durations remains to be seen. The company calls it "the world's first native, fully multimodal interactive world model."⚠️ Vendor claimSource: DeepTech source
FCC ban requires 65% of robot components to be domestically sourced, hitting US startups first · adjacent
The US Federal Communications Commission added "advanced robotics devices" to a banned-sale list last month; new models of foreign-made humanoid robots, quadrupeds, and even robot vacuums and lawn mowers entering the US market must be assembled in the US, with at least 65% by value of components sourced domestically, rising to 75% by 2029; already-sold devices and research imports not intended for sale are unaffected. The problem is that the domestic supply chain doesn't yet exist. Anto Patrex, founder of San Francisco-based CosmicBrain AI, told Rest of World, "Even Apple manufactures in China — how can you expect more complex robots to be made in America?" He set up an assembly facility in Canada earlier this year, but key components still come from China, meaning under the new rule they still can't be sold in the US either. Elizabeth Williams, founder of beauty-robot maker Gemma, called doing prototype iteration in the US "simply impossible." Michael Perry, head of commercial strategy at Houston-based Persona AI, put it this way: "You have to give both the carrot and the stick." The real cost of this policy won't be settled until the compliance milestones of 2027–2029.Source: Rest of World source
NVIDIA robotics executive travels to South Korea to visit LG, less than a week after MOU signing · adjacent
Less than a week after LG Group and NVIDIA signed a strategic memorandum of understanding, Madison Huang, NVIDIA's senior director overseeing Omniverse and robotics product marketing, is reportedly set to visit South Korea and head directly to LG Electronics' robotics facility. The MOU covers humanoid robots, AI factories, and automotive applications.⚠️ RumoredSource: DIGITIMES source
Faraday Future to pursue independent funding and listing for its robotics business, targeting 2,000 units shipped by year-end · humanoid
Jia Yueting announced that FF's robotics business will explore independent funding and a listing, with a "four-core, fully intelligent" commercialization push planned for Q3 and a year-end shipment target of 2,000 units. Meanwhile, the company's half-year report showed revenue down 99.93% year-on-year. Given the absence of any verifiable production line or delivery record, this counts as a strategic statement, categorized alongside the year-end shipment target as an unrealized plan.⚠️ PlannedSource: EEFocus source
Hardware & Supply Chain
· TI TCAN6062: Texas Instruments has launched what it calls the industry's first commercial CAN XL transceiver, supporting a per-frame payload of 2,048 bytes at 20 Mb/s, and cutting signal ringing in complex networks by up to 80% — directly targeting the bandwidth needs of industrial and humanoid robots for real-time limb position commands and high-volume sensor data; Arthur Mutter, chair of the CAN XL Special Interest Group, called it "a better and cheaper CAN FD that can also tunnel Ethernet" source
· Miniature planetary roller screws: NOUSBOT (Chinese robotics component maker) disclosed a screw with a 1.5-millimeter diameter, 5.5mm nut outer diameter, C5-grade precision, and 50-newton per-unit load capacity; its Suzhou production line runs at full capacity 24 hours a day, producing 3 million units annually and still facing shortages, with plans to reach 10 million units by year-end, cutting unit price from the thousand-RMB range down to the hundred-RMB range. Previously, China's miniature components relied heavily on imports from Switzerland's Rollvis, Germany's Bosch Rexroth, and Sweden's Ewellix. Worth noting: for high-end C0–C3 grades with lead accuracy under 5 microns, German and Swiss suppliers still hold over 70% market share, and there's a one-to-two-year durability validation gap between component qualification and mass-production adoption by whole-unit makers source
· Harmonic Drive (6324.T): the harmonic reducer leader posted April–June quarter revenue of JPY 16.68 billion, up 23.6% year-on-year, with operating profit of JPY 1.84 billion (versus just JPY 122 million in the same period last year); full-year operating profit guidance was raised to JPY 8.5 billion. Quarterly orders reached JPY 24.13 billion, up 55.7% year-on-year, driven by industrial robot orders from Japan and Europe. However, its China subsidiary moved in the opposite direction — revenue of JPY 1.01 billion, down 17.9% year-on-year, with operating profit down 33.2%, which the company attributed to a high base from Chinese domestic robot makers' demand last year and weak industrial robot demand this year source
· Zhaowei Machinery & Electronics (02692.HK): dexterous-hand revenue was not separately disclosed in its half-year report, and analysts are focused on its "revenue growth without profit growth" pattern and how the dexterous-hand business's actual contribution is being reported source
Top comments (0)