Today's Highlights
· Mandatory national standard for L3/L4 finalized — vehicles that don't meet it barred from market starting July 2027
· Dyna-2 pretrained on 1 million hours of human video, takes turnkey task from 0 to 90%
· World Labs clones a single real-robot demo into thousands of scenarios; a pure-simulation policy runs on real hardware for 1 hour straight
· Mech-Mind (Chinese robotic vision/AI startup) clears Hong Kong Stock Exchange hearing, gross margin up from 39.1% to 64.6% in two years
· Pony.ai teams with Uber to enter 5 European cities, plans to deploy over 2,000 robotaxis
· NVIDIA SONIC has Unitree's G1 kick open a trash bin lid with its foot, published in Science Robotics
· Qixian's unmanned coffee shop makes 202 cups in 1 hour, earns Guinness certification
· Unitree's IPO priced at a 219.23x P/E ratio, versus a 38.56x industry average over the same period
1. Research Progress
Dyna-2: 1 million hours of human video produces embodied AI's first scaling law · world-model
The question that keeps coming up in embodied AI is whether performance keeps improving as you add more data. Dyna Robotics's world-action model Dyna-2 offers the most direct answer yet: scaling pretraining data from 1,000 hours all the way to 1 million hours (roughly 170 years of human experience) produced monotonic improvement across evaluation metrics, with pretraining relying entirely on first-person human video and no robot data at all. The clearest illustration is a lock-box key-turning task — at pretraining volumes up to 100,000 hours, no checkpoint could turn the key; at 1 million hours, the reported success rate reached 90%. The company claims this demonstrates a real scaling law from human behavior to robot behavior; so far this conclusion rests only on the vendor's own disclosed experiments, with no public checkpoint or independent replication.
Dyna Robotics (US) · world-action model Dyna-2 · analysis: Orient Securities research report source
World Labs R2S2R: one real-robot demo cloned into a thousand trainable worlds · world-model
World Labs, founded by Fei-Fei Li, has released its first results from its Real-to-Sim-to-Real engine: it reconstructs the robot, sensors, environment, and a single task demonstration into an interactive virtual world, then perturbs lighting, object position and count, scene layout, physical properties like friction, and camera viewpoint one at a time — generating thousands of variants from a single real-robot demonstration. The key distinction is that the reconstruction isn't just visually convincing; the physical behavior has to match too — the team runs the same action sequence in parallel in simulation and reality, comparing observations, object motion, and outcomes frame by frame. The task list includes cable routing, threading a flexible wire end through a hole, and two-handed packing — precisely the kind of deformable-object task where simulation tends to break down. The company states that control policies trained purely in simulation transferred directly to multiple real-robot platforms, including Stanford's ALOHA, and on 4 additional platforms each ran continuously for 1 hour with zero human intervention, completing tasks such as coiling a power cord with both hands, precisely placing test tubes, and pulling a marker out of a dense pile. The underlying technology comes from SceniX, which World Labs acquired in July.
World Labs (US) · R2S2R engine tech blog · analysis: eGamers source
NVIDIA SONIC: 700 hours of human motion data teaches a humanoid to use its feet like hands · locomotion
Whole-body control for humanoids has long suffered from "good at walking but not jumping, good at climbing but not walking," with each motion type reinforcement-learned separately. NVIDIA's SONIC simplifies the goal to a single task — tracking human motion and mapping it onto the robot — then scales it up hard: over 700 hours and more than 100 million frames of motion-capture data, a model with up to 42 million parameters, trained on up to 128 GPUs for a cumulative 21,000 GPU-hours. A single policy accepts teleoperation signals, video, text, or even music as input, and generates running, jumping, crawling, and grasping, smoothly transitioning between speeds and postures. The most striking demo has a Unitree G1 equipped with SONIC following a person through a continuous sequence of actions: picking up a drink container by hand, stepping on a trash bin pedal with its foot to open the lid, then dropping the container in — here the legs aren't just for locomotion but act as a manipulator alongside the arms within the same task. The research was published in Science Robotics.
NVIDIA (US) · Science Robotics, published August 12 · analysis: 동아사이언스 source
4D-WAM: 3D trajectories used only during training, zero added inference cost · world-model
World-action models perceive 2D projections but must execute 3D interactions — a gap previously bridged by aligning 4D features directly into the model. The team's preliminary experiments found this actually creates optimization conflicts: the intermediate representations of a DiT-based WAM and Trace Anything's 4D representations are nearly uncorrelated or even negatively correlated across most layers. 4D-WAM instead aligns motion patterns rather than absolute values, using two complementary auxiliary supervision signals — Motion Alignment for "how things move" and Destination Alignment for "where they end up." The gains concentrate in out-of-distribution scenarios: under LIBERO-Plus camera perturbations, success rate rose from 27.89% to 45.15%, an average 8.8-point improvement across seven perturbation types; switching to the Lingbot-VA backbone raised RoboTwin randomized-scenario performance from 34.6% to 41.8%. After training, the teacher model and alignment modules are entirely removed, so inference parameter count, latency, and memory footprint see zero increase.
4D-WAM team · validated on ARX LIFT2 dual-arm real robot · analysis: QbitAI source
Xidian University's "electric eel" sensor identifies a material before it even touches it · perception
Electric eels locate prey in dark, murky water by sensing distortions in their own electric field. A team led by Zhang Weiqiang at Xidian University (Xi'an University of Electronic Science and Technology) has turned this principle into a non-contact proximity sensor: a specially treated fluoropolymer that behaves like a miniature electrostatic battery, holding its charge over long periods after being charged and maintaining a stable electric field around itself. Field changes caused by an approaching object can be used to infer its conductivity, dielectric properties, and geometry. Sensitivity is the strongest part of this work — it can resolve displacements as small as 50 microns (about half the diameter of a human hair), corresponding to a roughly 1-volt change, and in lab tests showed no noticeable signal decay after more than 10,000 approach-retreat cycles. It doesn't depend on lighting conditions, and works on plastic, glass, and wood in addition to metal — on a production line it could distinguish two visually identical parts made of different materials. The researchers are upfront about current limitations: it still requires fairly close range to respond, and extending its range and accuracy in complex environments is the next step.
Zhang Weiqiang et al. (Xidian University) · Advanced Materials · analysis: South China Morning Post source
2. Funding & Deals
Mech-Mind | Hong Kong IPO clears hearing | 2025 revenue RMB 389 million · embodied
Following its May 7 filing acknowledgment from the CSRC for a Hong Kong listing, the company published its post-hearing information pack via the HKEX disclosure system on August 16. Rather than building complete robots, it makes standardized eye-brain-hand components that let robots "see clearly, reason correctly, and act precisely." Its prospectus states global cumulative deployments of over 27,000 units across nearly 50 countries and regions, and claims the No. 1 global market position in AI+3D vision-guided general-purpose robotic components by 2025 revenue, with a 22.1% share. The financial trajectory is more notable than revenue itself: revenue rose from RMB 181 million to RMB 389 million between 2023 and 2025, while gross margin over the same period climbed from 39.1% to 64.6%; adjusted net loss narrowed from RMB 334 million to RMB 109 million, and narrowed further to RMB 33.47 million in Q1 2026 — already below that quarter's RMB 38.46 million in R&D spending. Overseas revenue grew from RMB 58.6 million to RMB 195.5 million, with the overseas share reaching 50.3% in 2025, surpassing the Chinese domestic share for the first time. Sales expense ratio fell from 103% to 43% over three years. Shareholders include Sequoia China, IDG Capital, Meituan, Qiming Venture Partners, and Source Code Capital.Source: FX168 Financial source
Qidian Zhikong | Angel round | ~RMB 100 million · hardware
Led by Yinfeng Capital, with participation from Zhangjiang Kechuang, Pudong Venture Capital, Fourier Intelligence (Chinese humanoid robot maker), and two listed companies. The company makes non-contact optical six-axis force sensors, bypassing the industry's mainstream strain-gauge approach — a route notoriously hard to manufacture, prone to fatigue and creep, subject to significant cross-axis interference, and costly. Six-axis force sensors are a key component for dexterous hands and joint force control in humanoids; over the past year, competitors such as Kunwei Technology and Bluepoint Contact Robotics have each raised nine-figure-RMB rounds, as capital pushes this component from "expensive and scarce" toward "commodity standard part" — the company states its goal is to bring unit price down to the thousand-yuan range, a target not yet achieved.Source: Sohu source
RealMan Robotics | A-share IPO tutoring filing | Sponsor: Guotai Junan · embodied
The company has submitted IPO tutoring filing materials to the Beijing Securities Regulatory Bureau, having completed its conversion to a joint-stock company in late May. Its core business is ultra-lightweight humanoid-style robotic arms: its seven-DOF humanoid wrist arm has a rated single-arm payload of 5 kg and a peak payload of 9 kg, forming the RM65, RM75, and RML63 product line. In May 2026 its Changzhou embodied-intelligence data lab platform went into production, deploying 150 units of its RealBOT wheeled humanoid robots to run reliability testing on the robot body and joint modules. Its investor roster includes Ecovacs (which led its Series A+ round in March 2023); it completed nearly RMB 500 million in strategic funding this March and achieved operating profitability in 2025. According to STAR Market Daily, roughly 30 robotics companies including Luoshi Robotics and Xgrids Robot remain in the IPO pipeline — whether upstream actuator makers can go public independently is now being tested case by case.Source: STAR Market Daily source
Unitree | STAR Market IPO pricing result | P/E of 219.23x · humanoid
Following an oversubscription of 8,288x and completed payment, the last set of pricing figures is in: the offer priced at a 219.23x P/E ratio, against an industry-average static P/E of just 38.56x over the same period — a premium of nearly 5x. Online retail investors forfeited subscriptions on 8,734 shares, worth RMB 1.3171 million, which were underwritten by the sponsor. Another controversy before listing involved rumors of gray-market ("dark pool") trading, with intermediaries reportedly quoting RMB 410-520 per share to buy new shares — a premium of over 170% above the RMB 150.80 offer price. Reporters sought confirmation from multiple investment institutions, private equity firms, and securities wealth-management desks, and most responded that this was essentially unfounded; the lead underwriter also did not confirm any large-scale over-the-counter trading. Founder Wang Xingxing publicly stated he does not want the market to over-speculate.⚠️ Compiled figuresSource: Sina Finance source
Primary market weekly | Disclosed funding total ~RMB 10.6 billion | up 25% week-over-week · adjacent
Per CLS Venture Watch, disclosed primary-market funding totaled roughly RMB 10.6 billion this week, up 25% week-over-week, recovering about half of last week's RMB 8.459 billion figure, which itself had plunged 73.17%. The single largest deal was Qianheyibang's Series B round of over RMB 2 billion.Source: Sina Finance source
3. Commercialization & Deployment
Pony.ai × Uber expand to 5 European cities, plan to deploy over 2,000 robotaxis · autonomy
The two companies announced an expanded partnership on August 14, building on existing commercial operations in Zagreb, Croatia, to expand coverage to 5 European cities including Zagreb, with fleet scale rising to over 2,000 vehicles; the remaining cities and rollout timing will be announced in phases, and Zagreb service will simultaneously integrate into the Uber app. The division of labor is central to this model: Pony.ai provides the L4 system, vehicles, and operational experience; Uber provides ride-hailing, payments, customer service, and its hybrid mobility network; day-to-day fleet management goes to local partners, and vehicle capital investment and asset ownership are adjusted based on market conditions — the company calls this a "co-built fleet" model. Its seventh-generation robotaxi is already production-ready to automotive-grade standards, and the company states it has achieved per-vehicle profitability in Guangzhou and Shenzhen; co-founder and CEO James Peng has previously said vehicle costs are roughly one-quarter to one-fifth of Waymo's. Under the company's plan, global operations will cover over 20 cities with a fleet exceeding 3,500 vehicles by 2026, meaning this European deployment alone would account for more than half of that expansion target.⚠️ Planned figuresSource: China Business Journal source
Qixian Coffee's world-first 24-hour unmanned store opens, sets Guinness record with 202 cups in 1 hour · embodied
The store opened August 16 at Galaxy SOHO in Beijing, and was certified the same day by Guinness World Records for "most coffees served by a robotic coffee shop in one hour," with a result of 202 cups — one cup roughly every 18 seconds. The entire process, from grinding, extraction, and mixing to dispensing and shelving, is completed independently by robots, with each cup taking under 30 seconds to prepare; customers order via self-service and can fine-tune sweetness and coffee strength from 1% to 100%. Behind the process is a combination of JD.com's AI large model and embodied AI, aimed at three cost drivers plaguing traditional coffee shops: unstable quality control, high labor costs, and limited operating hours. Prior unmanned-retail attempts have mostly stalled at single-location pilots; the actual advance here is running standardized output, record-level throughput, and 24-hour operation simultaneously. Whether a single store can be replicated across multiple locations remains to be seen.Source: Sina Tech source
GAC Huilun's GoMateMini nearly 50 units deployed across 7 projects, longest run 12 months · humanoid
Following a nine-figure-RMB funding round announced earlier this week, GAC Group's subsidiary Guangdong Huilun Technology disclosed deployment figures: its fourth-generation wheeled-legged humanoid GoMateMini has been deployed at nearly 50 units across 7 projects, with the longest continuous run reaching 12 months, and cumulative orders approaching RMB 10 million. The company was founded only in February 2026. At a time when humanoid robot progress is largely measured by unit counts, "12 months longest continuous run" is a more telling indicator than shipment volume of whether the hardware is actually staying deployed and working in the field.Source: WantChinaTimes source
Karsan's autonomous bus e-ATAK begins passenger service at the Netherlands' Efteling theme park · autonomy
Turkish bus manufacturer Karsan's full-size SAE L4 bus began carrying passengers on August 12 at the Netherlands' flagship theme park Efteling, running a roughly 6-kilometer route between the main entrance and the Wonderhotel. The ride is free and accessible, and is described as the first public-road autonomous bus operation in North Brabant province. The software comes from ADASTEC, the same stack used in its Rotterdam-The Hague Airport project. The two projects' licensing status should be distinguished: the airport line has held a commercial operating license from the Dutch vehicle authority RDW to carry paying passengers since July 2025, while this Efteling route runs, per the partner's own description, under a public-road testing permit for a roughly two-month pilot period, over a route mixing public roads with private zones shared with cyclists, pedestrians, and hotel traffic.Source: Unite.AI source
Einride integrates its L4 system into DAF's "X" series electric heavy trucks · autonomy
Autonomous logistics company Einride and PACCAR's DAF Trucks announced a partnership to integrate the Einride Driver L4 system into DAF's premium heavy-truck platform, covering everything from the 12-tonne XB Electric urban delivery model to the long-haul-oriented XG and XG+ Electric. Einride was previously the first company to successfully deploy L4 trucks on European public roads; this partnership fills its biggest gap — a mass-manufacturing partner, with PACCAR also the parent of Kenworth and Peterbilt. Neither company has announced a timeline or volume for the integrated vehicles.⚠️ Planned figuresSource: ZDNet China source
4. Industry News
China issues mandatory national standard for L3/L4 autonomous driving; non-compliant vehicles barred from market starting July 2027 · autonomy
The "Safety Requirements for Autonomous Driving Systems in Intelligent Connected Vehicles" (GB 44721-2026) was approved and issued on July 30 by China's State Administration for Market Regulation and the Standardization Administration, taking effect July 1, 2027, applying to Category M and N vehicles equipped with L3 or L4 systems (excluding automated parking). This is China's first mandatory national standard for L3/L4 — and the distinction in status matters: a recommended standard is voluntary compliance, while a mandatory national standard is a legally binding technical bar that any mass-production vehicle claiming L3 or L4 capability must clear in full going forward, or be barred from market. Four gaps left by the prior recommended standard are now closed: first, automakers must build safety-assurance capability across four dimensions — safety policy, risk management, safety guarantee, and safety improvement — spanning the full lifecycle from design and development through manufacturing and post-deployment, marking the first entry into the automotive industry of the "Safety Case" mechanism commonly used in aerospace internationally; second, system safety must reach at least "the level of a competent and attentive human driver performing the dynamic driving task"; third, L3 systems must come standard with driver takeover-capability monitoring, with clearly specified activation and exit processes and "ready/active/exiting" state indicators, and automakers must disclose capability boundaries via their websites and in-vehicle displays; fourth, a verification framework combining "enterprise capability assessment + safety-case review + confirmatory testing" is to be built, verified by third-party bodies. An Autohome article citing Professor Ji Xuehong of North China University of Technology notes that liability lines are also now clarified: accidents caused by the vehicle within its operational design domain are the automaker's responsibility, while failure to take over within 10 seconds of a takeover request falls to the driver. For challengers, the real barrier isn't lidar or compute — it's this safety-case management system, which takes time to build up.Source: Autohome source
MIIT: China's humanoid robot production topped 40,000 units in H1, expected to exceed 100,000 for the full year · humanoid
Data from China's Ministry of Industry and Information Technology shows humanoid robot production in China exceeded 40,000 units in the first half of 2026, with the full-year figure expected to top 100,000. This figure should be read separately from previously disclosed shipment numbers — global humanoid robot shipments over the same period totaled roughly 19,100 units (per SmartAnalyticsGlobal, previously reported), and the nearly two-fold gap between production and shipments reflects current inventory and work-in-progress scale. Over the same period, 116,000 new companies were registered in the humanoid robotics sector, up 9.5% year-over-year, out of 561,000 total new companies registered nationwide across China's "8 emerging industries + 9 future industries" categories.Source: WantChinaTimes source
LG signs MOU with NVIDIA to co-develop bipedal humanoid, targeting a Q1 2027 unveiling · humanoid
Per Barron's, the two companies signed a memorandum on August 13 to jointly develop LG's first bipedal humanoid robot, expected to use NVIDIA's Isaac GR00T foundation model, Jetson Thor compute platform, and Halos safety system, with LG contributing its own actuators, sensors, and batteries. The division of labor reveals what LG is actually betting on: NVIDIA supplies the "brain," while LG tests whether its own components can carry the "body" — LG Electronics handles appliances and robotics, LG Innotek handles sensing components, LG Energy Solution handles batteries, and LG CNS handles industrial software and logistics systems, a group structure that means LG doesn't need to source a robot body externally and simply layer on NVIDIA's software. A more conservative testbed is a wheeled robot: LG plans to test its CLOiD unit at a washing-machine factory in Tennessee this year.⚠️ Planned figuresSource: Startup Fortune source
Second World Humanoid Robot Games opens August 22 with 2,056 robots competing · humanoid
The event opens August 22 at the National Speed Skating Oval ("Ice Ribbon"), with an expected 666 teams and 2,056 robots competing — a 138% increase in team count over the first edition. Over 5 days of competition, events span 30 athletic competitions including ball sports and combat events, and 21 scenario-based events including industrial assembly and household services. The organizing committee has set up "robot homes" near each venue for pre-competition training and charging, and built an embodied-AI robot management platform that can manage robots across brands and models. The rules have also been made more demanding: the number of teams competing in free-form gymnastics rose from 3 last year to 18, with the number of required moves increasing from 9 to over 20, spanning 8 categories including static poses, handstands, support moves, and running/jumping moves; each team must select at least 5 categories and 10 moves for a self-choreographed routine, and stationary demonstrations have been replaced by mobile performances.Source: Beijing Daily source
Feagine unveils cross-embodiment foundation model Fi0, paired with three tendon-driven soft robotic arms · embodied
Having to retrain a model's intelligence from scratch every time the robot body changes is one of the most awkward dependencies in current embodied AI. Feagine's Fi0 treats the robot's physical structure and current state as input information for generating actions, using this to preserve cross-embodiment task knowledge, while releasing three tendon-driven soft robotic arms — A01, A02, and A03 — for systematic validation: A01 has a single flexible segment, 2 DOF, weighs 750 grams, with a 200-gram payload; A02 has two segments, 4 DOF, and a 400-gram payload; A03 has three segments, 6+1 DOF, a 50-centimeter arm length, and a 600-gram payload. The company also states that when the model encounters a task beyond its ability, a single human demonstration is enough — the model uses that demonstration as context during inference without updating its parameters. Soft continuum robotic arms continuously change shape during operation, making them one of the hardest cases for cross-embodiment generalization, and this claim awaits third-party verification.⚠️ Vendor claimSource: Interesting Engineering source
American Airlines bans humanoid robots from flights starting August 17 · adjacent
The ban covers all domestic and international, mainline and regional flights: humanoid and animal-shaped robots may not be placed in overhead bins, seats, or checked baggage, and anything found at check-in will be treated as an undeclared hazardous item. The airline cites design, size, and battery concerns, though no related incidents have been recorded so far; Southwest, Delta, and United already have similar rules. The problem lies in the thresholds: neither "small" nor "humanoid" is defined, leaving enforcement to front-line staff, and the policy hasn't yet been published on the airline's website — travelers carrying robot assistants could be turned away with no warning. The ban targets form factor, not actual battery risk, which is especially thorny for future assistive robots: federal accessibility regulations may eventually classify such devices as equipment rather than pets, yet the current policy excludes them across the board regardless of function.Source: InsideFlyer source
California Teamsters union sues DMV over new driverless heavy-truck rules · autonomy
California previously banned driverless heavy trucks until this May, when the DMV opened its autonomous-vehicle permitting system to heavy vehicles under what is now the strictest such framework in the US: operators must first complete at least 500,000 miles of testing with a safety driver and submit detailed data to the DMV; once approved, they must then complete at least another 500,000 miles of driverless testing, for a cumulative total of no less than 1 million miles of public-road testing (of which at least 200,000 miles must be in California) before applying for a commercial deployment permit. The California Teamsters union filed suit this month on procedural grounds, arguing primarily that the DMV failed to properly estimate the new rule's economic impact under the California Administrative Procedure Act. This is the first formal legal pushback from organized labor against the advance of autonomous trucking.Source: Orange County Register (opinion) source
GGII blue paper: value in intelligent welding is shifting from the robot body to perception and software · industrial
The GGII (High-Tech Industry Research Institute) released the "2026 China Intelligent Welding Robot Industry Development Blue Paper" jointly with Fastems, Elite Robot, and STEP Electric at its Intelligent Welding Robot Summit on August 12. GGII data shows global intelligent welding robot sales of 9,200 units in 2025, worth roughly RMB 2.162 billion, with China accounting for 6,900 units and RMB 1.38 billion — about 75% of global sales; over the same period, China's overall welding robot sales reached 59,500 units, up 13.33% year-over-year. The cost structure is even more telling: vision/laser tracking systems account for 26% of cost, intelligent welding control systems 9%, and offline programming software 5% — a combined 40% — while the robot body itself accounts for only 36.4%. That 40% is exactly the incremental market that distinguishes intelligent welding from traditional teach-and-repeat welding. Steel structures (China's 2025 national output was 107 million tonnes, requiring roughly 350,000 welders) and shipbuilding (welding accounts for roughly 30%-40% of total construction labor) are currently the two main large-scale entry points.Source: Sina Finance source
Hardware & Supply Chain
· Thundersoft's Tiangong 100: an automotive-grade 7nm AI accelerator chip, announced entering mass production on August 12, with volume supply now starting to automakers and Tier 1 suppliers source
· Touchlab electronic skin: uses quantum tunneling to measure pressure, force, and direction, with material thinner than human skin — aimed at the long-standing problem of a robot not knowing how much force to use when, say, gripping a strawberry; the company itself notes that breakthrough technologies typically enter the market at a high price first, with costs coming down later through scale source
Top comments (0)