Today's Highlights
· SoftBank in talks to acquire controlling stake in 1X, valuation $6 billion, five years after selling Boston Dynamics
· XPeng's second-generation VLA builds time into the model, remembering the past 30 seconds and projecting 6 seconds ahead
· XPeng Robotaxi secures Guangzhou remote-testing qualification, entering driverless front-seat road testing
· Waymo shares ten lessons from 200 million miles: L2-to-L4 upgrading is a "false summit"
· South Korea to invest 2.3 trillion won in full-stack humanoid robotics by 2030
· Tiangong Ultra clocks 8.64 seconds in the 100m final; AgiBot (Chinese humanoid robotics company) tops medal table with 46
· Yikong Zhijia (Chinese autonomous mining-truck company)'s driverless mining trucks double to 3,100 units, yet revenue falls 17.3%
Research Papers
One policy runs arms, humanoids and human hands: Xiaomi team's camera-centric action space · vla
VLA models aim to consume all robot data, but are stuck on incompatible embodiments, camera setups and underlying action spaces — the standard industry fix is a separate adaptation branch per dataset. Xiaomi's Embodied Intelligence team and the University of Macau took a different approach: instead of treating robot commands as the shared target, they redefine actions using anchor motion visible to the camera, so that arms, humanoids and human hands all become different carriers of the same action pattern, with a geometric-conditioning translator combined with the target embodiment's kinematics to recover executable control. Trained on 4.03K hours of robot and simulation data plus 2.34K hours of human demonstrations, a single checkpoint with no task-specific fine-tuning scored 98.3% on LIBERO, 82.0% zero-shot on LIBERO-Plus, and 89.2% on RoboTwin's hard scenarios.
Shaoqing Xu et al. (Xiaomi Embodied Intelligence Team · University of Macau) · arXiv 2608.26058 source
Zero-WAM: robots learn unseen tasks from a single human video used as a prompt · vla
Large language models can perform new tasks purely from context, but robots can't — the bottleneck is "task specification" itself. This work argues that the most natural specification for manipulation tasks isn't language but human video, since visual cues directly describe how the task should unfold. To address the scarcity of paired data, the team built an automatic pipeline converting robot trajectories into semantically matched human videos, assembling HumanGen (74.2K pairs covering 8.6K tasks), and used in-context future-clip prediction to discourage the model from taking shortcuts. Across seven unseen tasks in RoboTwin 2.0, average success rate reached 47.0%, 29.5 percentage points above the strongest video-action baseline.
Jiaming Zhou et al. · arXiv 2608.26103 source
R³: getting robots to reason through hard problems in natural language before acting · vla
Language reasoning has bought foundation models extra test-time compute, but whether that mechanism transfers to manipulation tasks was previously unclear. R³ is a two-stage post-training recipe: first mid-train an off-the-shelf VLM on expert reasoning trajectories to lock in a reasoning style, then apply single-step rubric RL using offline action data. Unlike prior work that treats structured trajectories as auxiliary supervision, it trains free-form language reasoning whose output directly guides a low-level policy. Across two testbeds — Language Table and a simulated dual-arm bagging task — both exploration and generalization on unseen tasks beat instruction-only imitation-learning baselines.
Lehong Wu et al. · arXiv 2608.26053 source
StreamPI: adding temporal memory to single-frame VLA without adding a single parameter · vla
Hugging Face trending score of 15 and rising. The strongest current VLAs, such as pi0.5, still operate frame-by-frame, with no memory of what just happened. StreamPI treats each (visual observation, language instruction) pair as an atomic time unit: bidirectional attention within a unit handles cross-modal fusion, while causal attention across units preserves autoregressive streaming inference, with the language instruction serving throughout as a semantic anchor. Training randomly samples frame intervals (e.g., every 3rd frame), yielding robustness to the frame-timing jitter of asynchronous real-robot deployment. On real-robot memory and fine-perception tasks plus LIBERO, it outperforms pi0.5 while directly inheriting single-frame pretrained weights.
Zhe Liu et al. · arXiv 2608.26067 source
TacForcing: touch can still change its mind mid-execution of an action chunk · manipulation
Contact state can change within a single action chunk, but chunk-based VLA tactile conditioning is frozen at the moment before execution — effectively working with stale information. Prior fixes bolted on a separate high-frequency reactive controller, complicating both architecture and training. TacForcing instead replaces the standard action expert with a streaming version, letting action generation track tactile observations arriving continuously during execution, and uses Execution-Aware Tactile Attention to apply tactile conditioning only to the steps nearest execution. Across six UniVTAC simulated tasks and three contact-rich real-robot tasks, average success rates reached 65% and 69% respectively.
Jianbo Zhou et al. · arXiv 2608.25798 source
MA-VLA: breaking cooperation into atomic instructions so multi-arm coordination can recombine · manipulation
Most VLAs treat language as a single global instruction with no mechanism to specify which arm does what, so they collapse on cooperation patterns absent from training. MA-VLA decomposes cooperative behavior into mid-level atomic prompts assigned to each arm, and introduces Arm Shuffle during training — permuting each arm's observations, state and atomic prompt — to force the model to execute instructions role-agnostically. The team also built a test set whose cooperation patterns are entirely absent from training, on which previous SOTA VLAs largely fail. Code, model and data are public; Hugging Face trending score of 5.
Zaibin Zhang et al. · arXiv 2608.25864 source
Twelve physical AI benchmarks turn out highly redundant, two pairs can simply be merged · benchmark
Model evaluation sets rarely overlap across papers, the model×benchmark matrix is sparse, and no one had measured whether benchmarks actually test the same thing. The authors selected 51 models × 12 benchmarks by reporting density from a registry of 51 benchmarks and 152 models, mixing model-card scores, benchmark-paper scores and their own reproductions, and found significant redundancy: merging two pairs of substitutable benchmarks into single columns shifted the equal-weighted average ranking of 22 of the 51 models by more than three positions. Greedy selection based on score dispersion plus variance unexplained by the already-selected set retained 78.5% of the utility of all 12 benchmarks using just four.
Zaruhi Navasardyan, Hrant Davtyan · arXiv 2608.25940 source
Super Odometry 2.0: holds pose through smoke, sandstorms and low light · perception
Odometry systems tend to collapse in smoke, sandstorms, snow or low light — precisely the daily conditions of search-and-rescue and mining. This version elevates the IMU to equal standing with cameras and LiDAR in sensor fusion, letting it take over when exteroceptive sensors fail, paired with a hierarchical adaptive mechanism for varying degrees of degradation. This is the most-cited paper in today's batch, with 12 citations.
Shibo Zhao et al. · arXiv 2608.25427 source
SkyDrive: watching traffic from drones overhead to teach autonomous driving new cities · autonomy
Planners trained via imitation learning often see large zero-shot performance drops in a new city, as traffic rules, road networks and driving habits all change — and re-collecting local data means deploying an entire vehicle sensor suite. SkyDrive instead uses drones for traffic-monitoring supervision: vehicle-mounted collection only records the ego vehicle and its surroundings, whereas a single aerial pass can observe large numbers of road users simultaneously, turning every vehicle into a data source. Drawing 650,000 driving samples from 137 hours of aerial footage and building a benchmark from it, experiments show that just 30 minutes of monitoring per location noticeably narrows the cross-city gap for most models. Data and code will be released.
Weijiang Xiong et al. · arXiv 2608.25142 source
Other papers today: GaussianDream++ adds metric 3D supervision to VLA via 3D Gaussian reconstruction and future prediction (arXiv 2608.25659 source); ConfAL-WM selects samples by confidence for post-training action-conditioned world models, improving efficiency on RoboTwin2.0 (arXiv 2608.25572 source); V-Link fixes VLA's action expert lacking access to visual representations (arXiv 2608.25308 source); VISTA infers spatial contact attention from vision, beating 3D Diffusion Policy on cap-twisting and calligraphy tasks (arXiv 2608.25872 source); LAC gives a humanoid upper body simultaneous linear and angular compliance (arXiv 2608.25405 source); RA-VLA does test-time adaptation via retrieval augmentation (arXiv 2608.25585 source); PRISM uses a GPU physics simulator as an online world model for dual-arm MPC (arXiv 2608.25666 source); 4DStreamCtrl unifies camera motion, object trajectories and depth into a 3D point-track representation (arXiv 2608.25479 source); CRESSim-Neo is a batched GPU simulation engine for surgical robots (arXiv 2608.25192 source); VBVR-Pro turns native visual reasoning into a verifiable closed-loop testbed (arXiv 2608.26105 source).
Open Source · Tools · Benchmarks
· Microduck: Hugging Face's second robot, priced at $399 and shipping this year, developed by French company Pollen Robotics (acquired last year). About 25cm tall and 770g, with LiDAR, camera, microphone and dual NFC readers, programmable in Python or JavaScript, and can be taught new skills via reinforcement learning and simulation. Its predecessor, Reachy Mini, has sold roughly 10,000 units to date. source
· DEEPRobotics (Chinese robotics company)'s DM0.5 tops RoboDojo: following this month's report that the top model on this real-robot leaderboard — launched by the University of Hong Kong and nearly 20 other institutions — scored just 12.8%, DM0.5 took first place with a composite score of 24.90 and average success rate of 19.34%, with its Memory-dimension score of 47.74 opening the gap; LIBERO composite success rate was 99.0%. The model is open-sourced. source
· NVIDIA COMPASS: a cross-embodiment navigation policy training framework that uses an AI agent to generate training curricula, aiming to help robots learn to navigate new environments faster. source
· South Korea opens 1.56 trillion tokens of AI training data: government-led dataset construction, made publicly available. source
Funding & Deals
1X Technologies (Norway) | Controlling-stake talks | Valuation ~$6 billion · humanoid ⚠️ Reported, unconfirmed
The Information, citing people familiar with the matter, reports that SoftBank is in talks to acquire a majority stake in this OpenAI-backed humanoid robotics company; neither side has confirmed it publicly. The timing is worth noting: just over a month ago, Hyundai announced it would buy SoftBank's remaining roughly 10% stake in Boston Dynamics, ending SoftBank's ties to Spot and Atlas entirely. When SoftBank sold an 80% controlling stake in Boston Dynamics to Hyundai in 2021, the whole company was valued at roughly $1.1 billion. Five years later, Masayoshi Son is now looking at a younger, earlier-stage-commercialization company priced at more than five times that. SoftBank's previous in-house robotics venture ended with Pepper, discontinued in 2020 after a cumulative production run of about 27,000 units.Sources: The Information source, 36Kr source
Ren Shaoqing's new company | NIO strategic investment | Unicorn-level valuation · embodied ⚠️ Single-party account
On August 24, NIO CEO William Li announced at an autonomous-driving all-hands meeting that Ren Shaoqing, NIO's head of intelligent driving, has founded an independent company for physical AI foundation models and embodied intelligence, with NIO making a strategic investment and entering a collaboration; Ren will continue heading NIO's intelligent-driving business. Ren Shaoqing is one of the four core authors of ResNet and a co-author of Faster R-CNN; in 2016 he co-founded Momenta, and in 2020 he joined NIO to build its autonomous-driving team from scratch. This "stay-in-role while founding a separate company" arrangement has no established precedent, and questions remain about how his time will be split under the dual role, and where the boundaries lie for vehicle data and intellectual property. The same day, XPeng's robotics unit announced a first round of over $900 million in funding (previously reported).Source: Robot Insight source
Psi Robot (灵初智能) | New round | Over $100 million · embodied
Investors include Tuopu Group, Chery Holding's Ruicheng Fund, Lens Technology, 37 Interactive Entertainment, Wuhu Municipal Investment Holding Group and Fosun Chuangfu, with existing investor Zhuhai Science & Technology Industry Group adding an above-allocation top-up. The company builds embodied "brain" models, with a proprietary dual-system architecture combining the Psi-R2 manipulation policy model and Psi-W0 action-conditioned world model — the former handling long-horizon task decomposition and continuous action planning, the latter forecasting and correcting action trajectories. Funds will go toward two areas: scaling up human-data collection, including building data-collection facilities across multiple locations; and pushing deployment in logistics and advanced manufacturing. Its previously reported automated optical-module production line, built jointly with YOFC, uses robots for sub-millimeter-precision insertion quality checks and vacuum sealing.Source: Gasgoo source
Yikong Zhijia (7687.HK) | First half-year report since listing | Revenue RMB 549 million · autonomy
The number of operating driverless mining trucks rose from over 1,600 a year earlier to over 3,100 (previously reported), yet revenue fell 17.3% year-on-year. The reason lies in the gross margin: the product mix shifted toward the lighter-asset "customer-provided fleet" model, whose share of revenue rose from 62.7% to 65.6%, pushing half-year gross profit up from RMB 11 million to RMB 104 million, a gross margin of 18.9%. The company posted a loss of RMB 189 million for the period, down from RMB 236 million a year earlier. It serves 38 mines, with 4 sites running fleets of over 200 trucks and the largest single-mine fleet at 566 units; over 1,000 additional trucks are on order and in the delivery pipeline. Cash and cash equivalents on hand stood at RMB 273 million as of June 30.Source: LeiDi source
Wonik Robotics (South Korea) | National Growth Fund investment | 350 billion won · industrial
Following this month's report of a planned raise of over 300 billion won, South Korea's National Growth Fund has formally finalized its investment, with the funds going toward building a new Jeonbuk AX Center. Also approved in the same batch was CJ 4DPLEX, with the two equity investments totaling roughly 570 billion won.Source: Seoul Economic Daily source
Motion (Brussels, Belgium) | Pre-seed | €1.7 million · humanoid
Offers "humanoids-as-a-service," supplying humanoid robots to European manufacturers on a service basis; the company says it's targeting Europe's robot-adoption gap rather than building its own hardware. Roughly $2 million.Source: EU-Startups source
Xijian Technology | Nearly RMB 500 million raised over the past year | Neuromorphic chips · hardware
Incubated out of Tsinghua University, builds neuromorphic chips.Source: ZDX (智东西) source
Commercialization & Deployment
LG Innotek mass-produces robotaxi camera modules for Zoox · hardware
5-megapixel modules, with different fields of view depending on their mounting position on the vehicle, together forming a 360-degree detection ring around the vehicle body; they're waterproof and must maintain performance in severe weather and sharp temperature swings. This marks LG Innotek's first mass-produced camera module built specifically for robotaxis; the two companies have collaborated since Zoox's early development, and the modules will be installed on vehicles over the coming years. Zoox completed a robotaxi factory in Hayward, California last year and is already running driverless ride-hailing in Las Vegas. LG Innotek is also working on LiDAR and radar, aiming to grow its autonomous-driving perception solutions business to 2 trillion won (about $1.45 billion) by 2030 and expand its customer base to automakers and robotaxi companies in North America, Europe and Japan.Sources: Businesskorea source, The Korea Times source
XPeng Robotaxi enters driverless front-seat road testing, completes 2,000 rides in internal trial within two months · autonomy
The Robotaxi, equipped with the second-generation VLA, has obtained remote-testing qualification for intelligent connected vehicles from the city of Guangzhou, allowing testing without a safety operator in the driver's seat on relevant grade-1, grade-2 and grade-3 test roads in Guangzhou. Since launching an internal employee trial in June, it has completed over 2,000 rides in two months, covering complex urban roads. The test fleet, based on the XPeng G9 platform, has been running in Guangzhou for five months since May. XPeng is pursuing a unified technical foundation, scaling the same base model up to L4 Robotaxi and distilling it down for production vehicles with lower on-board compute.Source: Kuaikeji source
Didi unveils Robotaxi R2; Beijing-Guangzhou demonstration zones now open to the public · autonomy
Co-developed with GAC Aion, the vehicle was formally unveiled at the 18th International Transport Exhibition. It carries 33 sensors, including 10 LiDAR units, cameras, 4D millimeter-wave radar, an infrared camera and sound sensors; GPU compute exceeds 2,000 TOPS, and Didi says its tri-domain-fusion central computing "brain" is the industry's first mass-produced solution of its kind. The vehicle is built on GAC's GEP 3.0 pure-electric platform, with electronic steering control and redundant braking. Portions of Huangpu district in Guangzhou have begun round-the-clock trial operation, with users hailing rides through the "autonomous driving" entry point in the Didi app; once it obtains a Beijing road-testing license, the R2 will join Beijing's "Huiju Port" operations network. The demonstration zones remain limited in scope, still far from full-area rollout.Source: WTOutlook (网通社) source
LimX Dynamics (Chinese humanoid startup)'s bipedal humanoid shipments pass 3,000 units · humanoid ⚠️ Company-reported
The company released operating figures on the closing day of the humanoid robot games: cumulative global shipments of its bipedal humanoids have surpassed 3,000 units, with over 2,000 shipped so far in 2026 and over 1,200 delivered in the first half of the year, putting it, by its own claim, among the top three globally. Shipments and deliveries are two different measures — the 3,000 and 2,000 figures are shipments, while the 1,200 figure is deliveries.Source: Guandian.cn source
Hyundai begins delivering IONIQ 5 autonomous vehicles to Waymo in Q4 · autonomy
The vehicles are produced at Hyundai's plant near Savannah, Georgia, then shipped to a Magna-operated retrofit facility to install the Waymo Driver, with a rooftop sensor suite including LiDAR, radar, cameras and microphones, plus additional sensors on the front and rear bumpers and front fenders. Waymo's previous-generation fleet is being replaced.Source: Sina Finance source
Turkey's Arvato deploys Lipbo (Chinese sorting-robot maker) 3D sorting robots, 3,700 packages per hour · industrial
Arvato, a Bertelsmann-owned logistics service provider, deployed the system in Turkey, with a processing capacity of 3,700 packages/hour.Source: Weidu.net source
Industry Developments
XPeng's second-generation VLA builds a time dimension into the model, rolling out in September · autonomy
It "remembers the world from the past 30 seconds" and projects scenarios 6 seconds into the future. At XPeng's Physical AI sharing day on the afternoon of August 27, the company named the capability behind its second-generation VLA Infini-VLA, built on a physical-world foundation model that for the first time incorporates a time dimension — officially described as AI's understanding of the world jumping from 3D space to 4D spacetime. Traditional inference is a serial loop of observe, compute, output, observe again; the new version switches to streaming autoregressive inference — observing, thinking and acting simultaneously — cutting end-to-end response time by 300%. The on-vehicle model's parameter count grew 3.5x, which the company says is more than 15 times that of mainstream industry VLA models, with multidimensional overall safety capability up 20x. In a live demo, the system remembered that a vehicle ahead was mid-U-turn, held back cautiously while waiting, and quickly moved off once the vehicle cleared. The new version rolls out in September to all Ultra and Ultra SE models, debuting first on the G9L; a distilled version, Turing VLA 2.0 Lite, will roll out the same month as a first batch to Max models with a single Turing chip. XPeng frames this version as reaching up to L4 while remaining compatible with more mass-production models, unifying L2 through L4 on a single technical foundation.Sources: Kuaikeji source, Guancha source
Waymo shares ten lessons from 200 million miles, aimed squarely at the L2-to-L4 upgrade path · autonomy
"Improving a driver-assistance system into full self-driving is a false summit. True L4 maturity can only come from safety built into the system by design, validated on closed courses, and hardened by the rigorous experience of driving with no one in the vehicle." Waymo's head of AI foundations, Srikanth Thirumalai, posted ten lessons on Tuesday, without once mentioning Tesla by name. The first lesson states that cameras alone cannot achieve safe autonomous driving at scale — Waymo uses cameras, LiDAR and radar as redundancy; the fourth targets pure end-to-end networks that map raw footage directly to steering commands: "you cannot build trust with a black box." Waymo currently provides over 500,000 paid fully driverless rides per week, targeting 1 million by year-end; Tesla confirmed last month that its unsupervised FSD had covered just 380,000 miles over a full year. The same day, XPeng announced it was unifying L2 through L4 on a single technical foundation — the two companies reaching opposite conclusions about the same technical path.Source: Electrek source
Humanoid robot games close: 100m in 8.64 seconds, AgiBot tops medal table with 46 · humanoid
Following this month's successive record-breaking times of 9.39 and 8.86 seconds, X-Humanoid's Tiangong Ultra brought the 100m down to 8.64 seconds in the closing-day final, 0.22 seconds faster than its repechage run, and nearly a second faster than Usain Bolt's human world record of 9.58 seconds. The fastest 100m time at last year's inaugural event was still over 12 seconds. Its closest rival in the final was Ubtech (Chinese humanoid robotics company)'s Lightning. Robots and humans compete under entirely different physiological, technical and rule conditions, so the times aren't directly comparable. AgiBot, competing for the first time, took 46 medals including 18 gold, topping both the gold-medal and overall-medal tables. X-Humanoid CEO Xiong Youjun said after the event that the "ChatGPT moment" for embodied intelligence is imminent.Sources: Interesting Engineering source, Robotics & Automation News source
Reuters feature: China's humanoid robots aren't yet smart enough to take your job · humanoid
Hardware leads the world, but intelligence can't yet support general-purpose work — that's the scene Reuters reporter Laurie Chen and colleagues observed at a training facility in Guangxi. The site has over 100 humanoid robots generating training data, moving slowly and clumsily; a novice trainer needs roughly 300 attempts to produce one usable motion, while a skilled trainer needs 50. UBTECH won an $18 million procurement contract from the Guangxi provincial government last October, supplying humanoid robots and related hardware for the site. Chinese humanoid shipments topped 40,000 units in the first half of the year, a 97% global share (previously reported) — but set against this usable-motion production rate, the lead in shipments and the gap in intelligence are two separate things.Source: Reuters source
AgiBot releases WALL-SS autoregressive world model · world-model ⚠️ Company-reported
Departing from diffusion models that generate video clip-by-clip, WALL-SS uses "next-scale autoregression" instead: sketching a coarse-scale outline first, then progressively refining image and physical detail. All three benchmark figures come from the company itself. Action-following score reached 0.29, which the company says is the only nonzero score among all models compared, with alignment between generated video and given trajectory commands rising from 0.251 to 0.539. Streaming rollout runs stably for 60 seconds without collapsing; by contrast, a model retaining only the most recent frames begins showing disappearing objects and warped tabletops after 20 to 30 seconds. The same policy showed a task-success-rate consistency of 0.93 between virtual and real-world settings, measured across six tasks and five training stages. The model has a built-in action-decoding head that reads the next action directly from a predicted frame, letting the same network serve as both simulator and planner.Source: Kuaikeji source
Teradyne sues JAKA's German subsidiary at the Unified Patent Court in Europe · industrial
The case was filed with the UPC's Copenhagen local division, involving software and hardware patents held by Teradyne Robotics A/S and its Universal Robots unit, targeting a range of collaborative robot models that JAKA sells in the EU. Teradyne Robotics president Jean-Pierre Hathout said, "we cannot and will not accept companies that illegally copy protected technology." This is the company's second IP lawsuit in Europe in 2026, following an earlier German suit over Universal Robots software copyright against another robotics company's subsidiary. Any ruling could apply across 17 of the 18 EU member states participating in the UPC, with the UK and Spain also named in this case.Source: Robotics & Automation News source
Hyundai considers selling Boston Dynamics robots through its car dealership network · humanoid ⚠️ Remarks at an event
Hyundai's CEO discussed the possibility of selling humanoid and quadruped robots through existing dealership channels, without giving a timeline or pricing. Also disclosed at the same event: Hyundai plans to deploy over 25,000 robots across its own factories by 2028.Sources: The Times of India source, CarsGuide source
Renesas opens Physical AI and robotics lab in Beijing · hardware
Focused on chip and system-level solution development for next-generation robotics.Source: Business Wire source
Hardware · Supply Chain
· AgiBot dexterous hands: the company says its critical-point series has shipped over 20,000 units cumulatively, holds the largest global market share, and has accumulated over 50,000 hours of real-world data; the OmniHand that won 7 golds at the games belongs to the same line. These figures are company-reported and have not been independently verified. source
· Zhongke Lingxi (Chinese robotics company) LG-603 industrial dexterous hand: global debut. The company says the prior-generation platform ran continuously for over a year at a Lili Motor factory in Lu'an, with 10,000 hours of zero-fault operation — roughly equivalent to 300,000 work cycles. The unit measures 203×82×40mm, weighs 510g, has a static load capacity of 25kg, a minimum grip diameter of 1-2mm, and a maximum overall grip force of 85N. A lighter variant, the LGX-601, weighs 400g with a 10kg static load capacity, and has opened a 30-day presale. The reference point is the 10,000-hour service-life threshold for dexterous-hand servo joints set by the Ministry of Industry and Information Technology's "Humanoid Robot and Embodied Intelligence Standards System (2026 Edition)": ≥10,000 hours. source
· Group standard for wheeled humanoid mobile chassis: jointly initiated by 14 companies, currently recruiting co-drafting participants. source
· Xiamen builds out dexterous-hand industry: a local industrial initiative, with products aimed at a range of fine-motor-skill scenarios. source
This Week's Watch
South Korea to invest 2.3 trillion won by 2030 in full-stack humanoid robotics · humanoid
South Korea's Ministry of Economy and Finance released this plan ahead of next year's budget announcement: 2.3 trillion won (about $1.66 billion) to be invested by 2030, with 600 billion won allocated for next year alone. The targets are specific: build a mass-production system for humanoid robots for 10 major industries; raise the localization rate of key components from the current 45% to 80%; procure 250 domestically made humanoid robots in 2027 for distribution to universities and state-funded research institutions, reaching a cumulative 1,080 humanoids by 2030, plus roughly 5,000 total including quadrupeds and other non-humanoid platforms. R&D focus is set on three components where South Korea's competitiveness lags — actuators, robotic hands and sensors — plus high-performance AI chips and batteries for humanoid robots. The government says it aims to complete full-stack humanoid technology, including a proprietary robot operating system, by 2031.Source: Korea JoongAng Daily source
US adds foreign-made advanced robots to FCC covered list · industrial
This month's previously reported domestic-component-ratio requirement now has its full rule text. The FCC Public Safety and Homeland Security Bureau's designation took effect on July 28, 2026, adding humanoids, quadrupeds, AMRs and other wheeled or tracked ground platforms as a category to the covered list under the Secure and Trusted Communications Networks Act. Being listed means the products cannot obtain FCC equipment certification — effectively barring import, marketing or sale in the United States — unless granted conditional approval by the Department of Defense, with an application deadline of January 1, 2028. To avoid this classification, a robot must be manufactured in the US and meet a domestic-component cost threshold, which is 65% through 2028 and rises to 75% in 2029. Models that already received FCC certification before July 28, 2026 are unaffected. ARC Advisory published an analysis this week on the practical impact for industrial robots.Sources: FCC official FAQ, Morgan Lewis, ARC Advisory source
Three market forecasts, three different bases: humanoid sales, robotaxi fleets, cost-parity point · adjacent ⚠️ Compiled from multiple sources
Robot Insight cites a Morgan Stanley forecast that Chinese humanoid robot sales will grow 133% year-on-year to 28,000 units in 2026, potentially reaching 2.6 million by 2035. The same article cites a Soochow Securities research note giving a cost threshold: once a robot's unit price falls below RMB 400,000, replacing human labor becomes economically viable. On the robotaxi side, Businesskorea cites Counterpoint Research data projecting a global robotaxi fleet of 3.6 million vehicles by 2035, with a service market worth $168 billion. The three figures come from three different organizations using three different methodologies and don't corroborate one another. Soochow Securities separately published an in-depth computing-industry report this week on the re-rating potential for industrial software and automation as world models shift from "generating the world" to "controlling the world."Sources: Robot Insight source, Businesskorea source, Faxian Baogao source
This week in supply chains: dexterous hands move from the arena to the production line, service-life becomes the new bar · hardware
The most concentrated supply-chain activity this week centered on end effectors. The robot games featured a dedicated dexterous-hand event, with a group of manufacturers including AgiBot, LimX Dynamics and Booster Robotics (Chinese humanoid startup) using it as a marketing showcase; at the WRC show, 373 companies and 311 new products debuted, with component makers unusually taking center stage — audience questions shifted from "what motions can it do" to "how long can it run reliably." The line that actually got drawn came from the Ministry of Industry and Information Technology's "Humanoid Robot and Embodied Intelligence Standards System (2026 Edition)," which sets a dexterous-hand servo-joint service-life requirement of ≥10,000 hours. Zhongke Lingxi's claim of 10,000 hours zero-fault operation over a full year at the Lu'an Lili Motor plant, and AgiBot's reported cumulative shipment of 20,000 units, both benchmark against this line — but both are self-reported figures with no third-party verification available yet. Pressure is coming from both directions: integrators developing end effectors in-house, cross-industry entrants flooding in, and capacity expanding too fast, squeezing component suppliers in between. South Korea's 2.3-trillion-won plan naming actuators, robotic hands and sensors as its three weak points also suggests that the difficulty of replacing these three components sits at roughly the same order of magnitude across countries. No material price or capacity changes elsewhere upstream to report this week.Sources: China.com source, NetEase source, Korea JoongAng Daily source
