Today's Highlights
· Daimon Robotics (Chinese tactile-sensing robotics startup) raises hundreds of millions of yuan led by Ant Group, Ant's first bet on robotic touch
· Global humanoid shipments hit 19,100 units in H1, with Chinese makers taking 97%
· Zoox launches in Las Vegas, the first paid steering-wheel-free robotaxi in the US
· Hesai Autonomous Driving's H1 revenue reaches 804 million yuan, with over 100 unmanned mining trucks now running routinely at 7 mine sites
· Linghou Robotics raises over 100 million yuan Series C, led by CATL-affiliated Chendao Capital
· Guao Technology raises nearly 100 million yuan Series A+, its micron-precision industrial robots now sold into Bosch's production lines
· RynnValue switches to time-to-goal as its supervision signal, real-robot success rate rises from 52.5% to 72.5%
· Tianjin sets a 2028 target of 20 billion yuan core output value for intelligent robotics
I. Research Progress
RynnValue: using "how much longer to the goal" as the supervision signal, turning the reward model into a foundation model · manipulation
General-purpose reward models are becoming the bottleneck that keeps robot learning from scaling — both preference labels and normalized-progress anchors transfer poorly across embodiments and data sources. RynnValue instead supervises on temporal distance (the directed cost-to-go from the current observation to a language-specified goal), with labels derivable directly from timestamps, letting it scale to over 7,000 hours and roughly 3 million instruction-conditioned segments without any preference or progress annotation. It achieves a Kendall's tau_a of 0.675 on RBM-EVAL-OOD, beating the SOTA under full preference supervision (0.655) and more than doubling the pure-progress baseline (0.292); converted into a dense reward, it lifts real-robot policy success rate online from 52.5% to 72.5%, and offline from 63.8% to 82.5%. The model is open-sourced and is today's top-trending paper on HF (HF↑6).
Dongchi Huang et al. · arXiv 2608.09853 source
SpeedTuning: imitation-learning policies run slow because they're locked to the demonstrator's hand speed · manipulation
Chelsea Finn's team flags a long-avoided deployment issue: imitation-learning policies are inherently capped by the speed of the hardware and the human who collected the data, and almost no one has measured the empirical curve of "how much success rate you lose per unit of speed-up." SpeedTuning uses a lightweight reinforcement-learning framework to predict the optimal execution speed for each action, requiring no extra data collection and layering purely on top of a base policy, achieving over 2.4x speed-up on dynamic and high-precision tasks like pouring, throwing, and grasping — while still outperforming naive acceleration methods such as fixed-rate linear interpolation.
David D. Yuan et al. (Stanford) · arXiv 2608.09138 source
HarnessWAM: World Action Models can only "predict the next segment" — bolting on a planning-capable outer brain · world-model
arXiv saw a burst of eight papers today around World Action Models (WAM, models that jointly model future frames and actions), spanning 4D trajectory-field injection, video-diffusion prior distillation, joint-embedding modeling, and speculative-decoding acceleration — this paradigm's density is now approaching last year's level for VLA. HarnessWAM targets one of its most concrete shortcomings: WAMs can only predict over a limited time horizon and lack global planning, cross-stage state persistence, execution verification, and failure recovery. It bolts on a VLM task manager to maintain scene beliefs and a structured task graph, using an event-driven dual-timescale loop to let the robot recover state after a subtask fails without discarding existing scene knowledge, achieving 59.6% full-task and 69.9% subtask success on RoboMemArena.
Zhaopeng Gu et al. · arXiv 2608.09516 source
SLIM-0.5B: continuous manipulation doesn't actually need open-domain semantics — 0.5B parameters suffice · vla
VLA policies spend most of their parameters supporting open-domain semantics, when continuous manipulation really only needs a compact representation of observation, action, and the state change the action causes. SLIM uses self-supervised masked trajectory prediction to learn action-grounded predictive latents, paired with a Mixture-of-Transformers backbone and flow-matching action generation, matching or beating large-scale VLA and world-action baselines on simulation benchmarks and real-robot evaluations with just 0.5B parameters — with no additional embodiment-specific pretraining and significantly lower inference latency and memory footprint.
Jingkai Wang et al. · arXiv 2608.09771 source
MetaSpace: "metamorphic testing" for embodied agents catches over 90,000 spatial-cognition errors in one pass · benchmark
Existing embodied benchmarks either rely on manually labeled VQA or only look at task completion rate, and the latter lets agents "game" the metric via suboptimal or even unsafe paths, masking real flaws. MetaSpace borrows metamorphic testing from software engineering, automatically generating test cases from the spatiotemporal multimodal state of real execution trajectories, encoding metamorphic relations grounded in logical rules and physical laws as executable Prolog rules — any violation is flagged as a spatial-cognition failure. It detects 90,422 spatial-cognition errors across three classes of embodied scenarios; every SOTA multimodal large-model agent tested scores only 0.44–0.52 on spatial cognition, versus a human baseline of 0.96.
Gengyang Xu et al. · arXiv 2608.07533 source
Training a VLM into a VLA causes depth perception to collapse in the final layers · vla
A counterintuitive negative result: the authors take a weight-aligned open-source VLM/VLA pair (Molmo2-ER and MolmoAct2-LIBERO) and probe layer-by-layer depth decodability, finding the VLA is worse than the original VLM at every layer (which they call the floor), and the degradation is uneven — the base VLM's depth decodability keeps rising in the final layers, while the VLA collapses outright (which they call the cliff). They causally trace this cliff to interference from the final-layer MLP writes: ablating the final-layer MLP write recovers most of the decodability, while equivalent attention ablation, or the same intervention on the base VLM, has no such effect. This implies action post-training is quietly eating away at the vision-language backbone's spatial-geometric capability.
Alexander Hackett et al. · arXiv 2608.08904 source
LUCID: long-horizon humanoid loco-manipulation planned at the high level with "imagined dynamics" · locomotion
Long-horizon humanoid mobile manipulation has mostly relied on scripted planners or finite-state machines to sequence pretrained skills, which breaks down as task sequences grow complex. LUCID first trains a low-level whole-body policy conditioned on structured latent variables via adversarial imitation and freezes it, then jointly learns a high-level policy and a macro-dynamics world model — the latter predicts the temporally extended state transitions induced by latent-variable decisions, letting the high-level policy optimize over imagined rollouts. On multi-object rearrangement tasks, it outperforms existing baselines on both full-task success rate and partial-completion rate.
Cheng Guo et al. · arXiv 2608.07746 source
Other papers today: 4D-WAM injects spatiotemporal awareness into world action models via 3D trajectory fields, a model-agnostic training strategy (arXiv 2608.08023 source); Vid2WAM offline-distills diffusion priors from large video foundation models into a compact WAM, discarding the teacher at inference (arXiv 2608.08558 source); JEPA-WAM learns VLA policies via joint-embedding world modeling, avoiding the overhead of explicit future generation (arXiv 2608.09381 source); SG-WAM adds text-grounding and spatial-awareness semantic guidance to WAMs (arXiv 2608.08839 source); WA-SpecDec applies world-aware speculative decoding to VLA, compressing closed-loop latency (arXiv 2608.08725 source); WorldSimProbe diagnoses "simulation fidelity" in action-conditioned world models, distinguishing accurate transitions from merely plausible ones (arXiv 2608.09298 source); Auditing Instruction-Trajectory Mismatches performs training-free audits of demonstration data, catching samples where the trajectory is correct but the paired language instruction is wrong (arXiv 2608.07895 source); AdaDexGrasp fuses 3D visual and tactile representations for post-contact adaptive grasping (arXiv 2608.07600 source); an ultra-low-impedance gripper drops the high-reduction-ratio transmission and external force sensor to enable high-bandwidth transparent physical interaction (arXiv 2608.09198 source); PhysX-CoT rewrites single-image-to-simulatable-3D-asset generation into an explicit, structured physical-reasoning chain (arXiv 2608.08053 source); CMU-Drive and V2V-VLA are closed-loop reasoning benchmarks for multi-vehicle cooperative driving (arXiv 2608.07621 source); Explore, Map, Remember, Decide asks whether embodied VLMs' spatial memory holds up in safety-critical scenarios (arXiv 2608.08077 source).
Open Source · Tools · Benchmarks
· XPolicyLab: cuts the O(NM) integration cost of connecting N policies to M evaluation environments down to O(N+M) via a unified standard and open ecosystem, already incorporating 42 robot policies; onboarding time for one representative policy drops from over 5 hours to 2 hours, and to 30 minutes with a packaged agent skill; the same adapter suite serves RoboTwin, RoboDojo simulation, and standardized real-robot evaluation (arXiv 2608.09892 source)
· LTX-2.5: LTX releases a new open-weight world model, using a new pixel-diffusion pipeline to reduce visual artifacts and improve cross-generation consistency, optimized for local inference on NVIDIA RTX GPUs, downloadable now on HuggingFace and ComfyUI; the LTX model family has surpassed 33 million cumulative downloads on HuggingFace source
· Daimon's open-sourced tactile dataset: the first open-source batch, a 10,000-hour all-modality manipulation dataset including tactile data, has surpassed 4.4 million downloads, ranking first among embodied-AI datasets on ModelScope; also released alongside is RobOmni, a matching all-modality evaluation benchmark including tactile data source
· EsaacSim: a multimodal event-camera plugin for NVIDIA Isaac Sim, filling a long-standing gap in event-vision tooling for mainstream robot simulation platforms (arXiv 2608.08522 source)
· Ego-OSCAR: an open-hardware head-mounted stereo-inertial capture rig, with hardware-synchronized global-shutter stereo cameras + 6-axis IMU + on-board encoding on an embedded Linux board, built for in-the-wild first-person data collection (arXiv 2608.08285 source)
· 360CityArena: an urban navigation benchmark built from real 360-degree video, filling a gap left by existing outdoor benchmarks that are either not realistic enough or not complex enough (arXiv 2608.08814 source)
II. Funding & Deals
Daimon Robotics | Strategic funding | Hundreds of millions of yuan · embodied
Led by Ant Group, with existing shareholders oversubscribing to follow on. This marks Ant's first move into tactile sensing as a physical-interaction layer, after previous bets on Unitree, Galbot, and AgiBot Intelligence (Chinese humanoid/embodied startups). Daimon spun out of a Hong Kong University of Science and Technology team and began commercializing in 2023, closing a hundred-million-yuan Series A and this new round within two months; its shareholder base now spans Inovance's investment arm, China Mobile's supply-chain fund, Legend Capital, China Merchants Capital, and China Telecom. Co-founder and chief scientist Wang Yu was the first PhD graduate of Matthew T. Mason, known as "the father of robotic dexterous manipulation." The same day, the company released Daimon-TWM, which it describes as the world's first tactile-anchored world model, unifying physical cognition, deliberative decision-making, and instantaneous manipulation into a single framework; the company says it has already deployed the system in 3C precision manufacturing and automotive production lines for "feel"-dependent tasks such as USB plug/unplug and fragile-item sorting, with a pencil-case-organizing demo at WAIC achieving a 95% on-site success rate. ⚠️ Vendor-reported figures The hardware-side numbers are harder to verify independently: the company claims to be the world's first vision-tactile sensor maker to reach 10,000-unit-scale mass production and delivery, with over 200 cumulative customers (more than 50 overseas), having already shipped devices to OpenAI, Figure, Physical Intelligence, Meta, BMW, and Google DeepMind. Its home-data-collection network built jointly with China Mobile has deployed 1,000 units in its first phase, with projected annual output of over 1 million hours of real-scenario manipulation data at full capacity.Source: SmartHey source
Linghou Robotics | Series C | Over 100 million yuan · hardware
Led by Chendao Capital, an industrial investment platform affiliated with CATL, with follow-on from an industrial-machine-tool investment fund, Yuxin Electric Technology's affiliated fund, and a dual-currency fund, plus additional investment from existing shareholders Suzhou Venture Group and Guyu Capital. This core-component maker, founded in Suzhou in 2015, has completed five prior funding rounds — including from NIO — with cumulative funding exceeding 500 million yuan; it now has over 700 employees (300+ in R&D), serves over 900 customers in China, and its product line covers machine-vision systems, direct-drive motor systems, industrial robots, and general-purpose robots. Of the three stated uses of the new funds, the most notable is the third — building Jiangsu province's first provincial key laboratory for embodied-AI robot core technology and components, pushing the company from a component supplier toward setting standards and shared common technologies.Source: 智东西 source
Guao Technology | Series A+ | Nearly 100 million yuan · hardware
Investment from Shenzhen Capital Group, Cornerstone Capital, and Guangdong Technology Financial Investment; Shenzhen Capital Group is an existing shareholder from last June's Series A. Funds are earmarked for mass production of industrial embodied robots and building out a robotics business unit. This company, founded in 2019, makes "multi-degree-of-freedom motors" — using a single mover structure to achieve both linear and rotary motion simultaneously via magnetic-field changes, fully direct-drive with no mechanical transmission; its ZR motors are already in the supply chains of Xinyichang, Changchuan Technology, Goertek, and Sunny Optical. Industrial robots built on this motor have already been sold to Bosch in Germany and are in pre-production trials on several major manufacturers' lines; its G-Tools, released this April, can grasp a 0.5mm-diameter fine needle through a micro-hole with ±2μm positioning accuracy and ±0.05N force control. A comment from founder Li Siyang is worth noting: "3C manufacturers basically won't consider using general-purpose humanoid robots" — the gap isn't that the foundation models aren't strong enough, but that base-body motion control can't meet industrial requirements — "the benchmark isn't Optimus, it's ABB." The company's Ganzhou plant currently has capacity of 1,800–2,000 units/month, and it projects revenue of nearly 100 million yuan in 2026 with profitability.Source: 36氪 source
LatentVerse (Chinese world-model startup) | Seed round | Hundreds of millions of yuan · world-model
Led and followed by Hillhouse Ventures, Clearwater Capital, and Innoangel Fund, with AgiBot and RobotEra (Chinese humanoid-robot makers) both taking stakes. The company was only founded in May 2026, spun out of Tsinghua's Institute for Interdisciplinary Information Sciences; founder and CEO Hu Yucheng studied under RobotEra founder Chen Jianyu and previously worked on embodied foundation models at ByteDance Seed. The team's earlier work — DP3, HiRT, VPP — includes VPP, which has been cited by Nvidia's DreamZero and other world-action-model projects. Its in-development model, UTAM, is a unified foundation model combining a vision-language expert, a world-model expert, an action expert, and a high-frequency tactile expert — the action expert generates a Latent Action, which the tactile expert then reconstructs into a closed-loop executable action based on contact force, slip, and deformation.Source: 机器人前瞻 source
Huilun Technology (incubated within GAC) | New funding round | 100 million yuan · humanoid
This embodied-AI robotics company, incubated internally by Guangzhou Automobile Group (GAC) this February, has completed a 100-million-yuan funding round, headquartered in Guangzhou's Huangpu district. Huilun has taken over all of GAC's business in this area and operates as an independent, market-facing entity, with full in-house capability across structural design, systems platform, motion control, and perception/decision algorithms; it has developed four generations of products to date, and plans call for small-batch trial production of its core product this year, with mass production targeted for 2027.Source: 金融界 source
NdotLight (South Korea) | Follow-on investment | KRW 15 billion · adjacent
Naver's early-stage investment arm D2SF has made its third investment in this 3D synthetic-data company for robot training, with this round sized at KRW 15 billion. South Korea has recently sharply increased its focus on the data side of physical AI, filling in a gap after earlier moves in humanoid hardware and safety certification.Source: Chosunbiz source
III. Commercialization & Deployment
Zoox launches the US's first paid steering-wheel-free robotaxi in Las Vegas · autonomy
Amazon-owned Zoox began paid passenger rides in Las Vegas on August 10, the company's first-ever paid operation, having previously offered only free rides to select riders in San Francisco. Following last week's disclosure that it would begin paid rides "starting next week" with approval to deploy 5,000 vehicles within two years, this step turns the plan into actual revenue. The differentiator versus Waymo, which operates in the same city, is the vehicle itself: no steering wheel or accelerator pedal, built instead with a face-to-face carriage-style cabin, and no formal front or rear — it can drive bidirectionally, with passengers adjusting temperature and music via an in-cabin screen. Notably, around the same time as the launch, Zoox's CEO publicly said "we need to be regulated" after one of its robotaxis drove into a fire scene in Las Vegas — an unusual stance for an autonomous-vehicle company to take.Source: Benzinga source · Fox Business source
Hesai Autonomous Driving's H1 revenue nearly doubles to 804 million yuan; over 100 unmanned mining trucks now running routinely at 7 mine sites · autonomy
Following yesterday's disclosure of cumulative unmanned mining-truck shipments exceeding 1,900 units, up 213% year-on-year in H1, the interim results fill in the financial side of this deployment curve: H1 revenue reached 804 million yuan, up 96.98% year-on-year; gross profit reached 212 million yuan, up 204%; gross margin rose from 17.1% to 26.4%; the period loss narrowed 75.23% year-on-year to 113 million yuan, with adjusted net loss down to 19.68 million yuan. The autonomous-driving business contributed 784 million yuan in revenue, 97.6% of total revenue. The scale of deployment is the hardest data point in the report — over 1,900 mining trucks cover nearly 40 mine sites globally, with 7 sites now running over 100 trucks routinely, and one single site running over 220; the 500-vehicle purchase order signed with Guangna Group this February is, as of the reporting period, the largest single formal procurement order in the global unmanned mining-truck sector. Some financial items still bear watching: total borrowings rose to 687 million yuan, the capital-to-liability ratio rose from 29.4% to 41.1%, and the company flagged financial-guarantee and loan-impairment risk related to "Customer K," with impairment provisions of 81.6 million yuan already booked.Source: 集微网 source
Kodiak × Atlas Energy: unmanned heavy-truck fleet expanding from 28 to 100 trucks, public roads early next year · autonomy
Atlas Energy Solutions plans to expand its Kodiak-powered unmanned heavy-truck fleet — which hauls sand to oil and gas well sites in the Permian Basin — from 28 to 100 trucks by mid-2027. The trucks currently run on private industrial roads and are expected to begin operating on public roads in early 2027, though the timeline still depends on operational milestones.⚠️ Company plan, not yet realized What has already changed is the operating model: Atlas has moved from single-loading-point operations to dispatching simultaneously from two yards roughly a 90-minute drive apart in Texas and New Mexico, connected via its 42-mile Dune Express conveyor-belt system for the final leg. Note a discrepancy in the figures — the 28-truck figure here refers to Atlas's own fleet alone, and is not the same statistic as the overall Kodiak fleet size reported last week.Source: Fox News source
Agility's Digit V5 set for December delivery to early customers, pitched as "no fencing required" · humanoid
Agility Robotics will deploy its next-generation Digit V5 to early customers in December; the selling point is not new motions but an upgraded safety system — the robot can work in close proximity to workers without traditional safety fencing, saving on floor space, facility modification, and workflow-redesign costs, which is exactly the most practical barrier to humanoids entering warehouses and factories. The company says the model is designed to support a 20-hour workday and swappable end-effectors, with autonomous tool-changing left for a later stage.⚠️ Company plan, not yet realized Agility previously disclosed plans to go public via a $2.5 billion SPAC, raising over $620 million.Source: Interesting Engineering source
Hong Kong approves "air-rail interchange" autonomous-driving trial in West Kowloon, 4 private cars with backup operators · autonomy
Hong Kong's Transport Department announced on August 11 that it has approved, under the Road Traffic Ordinance, an "air-rail interchange" autonomous-vehicle trial running between the High Speed Rail's Hong Kong West Kowloon station and the Airport Express Kowloon station, involving 4 autonomously operated private cars. A backup operator will remain on board throughout the trial and take over if necessary. The scale is small, but it connects autonomous driving to the transfer route between a cross-border high-speed rail line and the airport express — a new scenario beyond Hong Kong's earlier minibus road trials.Source: 观点网 source
CJ OliveNetworks builds AI logistics system for HD Hyundai Electric's Cheongju plant · industrial
CJ OliveNetworks will build a next-generation smart logistics platform at HD Hyundai Electric's Cheongju site, bringing AI-driven logistics scheduling to heavy-electrical-equipment manufacturing. It's a typical path for South Korean conglomerates: an in-group IT subsidiary taking on smart-logistics upgrades for a sister manufacturing company.Source: The Korea Times source
Tate deploys 58 Hirebotics collaborative welding robots across multiple plants · industrial
Tate, a maker of data-center flooring and airflow-management equipment, has deployed 58 Hirebotics collaborative welding robots across its plants. Welding is one of the few processes where cobots have already scaled from single-site demos to multi-plant replication — deploying dozens of units at once says more about process standardization than a single pilot site does.Source: The Robot Report source
Tencent's ride-hailing service formally integrates WeRide's (Chinese autonomous-driving company) Robotaxi · autonomy
WeRide's robotaxi service is now integrated into Tencent's ride-hailing entry point, part of a push to expand consumer-facing distribution channels. What China's robotaxi operators generally lack isn't vehicles but traffic — integrating into a super-app is another aggregation move following earlier integrations with Amap and the WeChat ecosystem.Source: 36氪 source
IV. Industry Developments
Global humanoid shipments hit 19,100 units in H1, Chinese makers take 97%, but researchers warn hardware is outpacing the underlying technology · humanoid
Following yesterday's report that AgiBot overtook Unitree with roughly 8,400 units for 44% global share, this report fills in a more complete picture: citing data from San Francisco-based research firm Smart Analytics Global (SAG), Southern Metropolis Daily reports global humanoid-robot shipments reached about 19,100 units in H1 2026, up 272% year-on-year from 5,100 in H1 2025, with Chinese manufacturers accounting for over 97% of global shipments and China itself accounting for over 85% of global demand. Unitree grew 170% year-on-year to about 5,900 units, ranking second with 31% share; its flagship G1 model has shipped roughly 11,000 cumulative units as of mid-2026. Behind the top two, Galbot, UBTECH, and AgileX shipped roughly 900, 700, and 600 units respectively in H1 — a clear gap from the leaders. The structural shift matters more than the total: industrial and commercial applications now account for over 70% of global shipments, up from about 50% in H1 2025, with manufacturing and warehousing/logistics becoming the main deployment channels. SAG projects nearly 60,000 units shipped for full-year 2026, with industry revenue of about $1.6 billion. But the report states plainly that "hardware scale is growing faster than the underlying technology is maturing" — real training data, model capability, reliability, manipulation precision, and cost remain constraints on broader deployment; separately, the FCC added "advanced robotics equipment" to its regulated list on July 28, which could create headwinds for Chinese suppliers' expansion into mature overseas markets.Source: 南方都市报 source
Dexterous hands raise over 25 billion yuan in H1 alone, and the industry still hasn't converged on a "standard hand" · embodied
According to incomplete statistics from IT Juzi, China's dexterous-hand sector raised about 16.877 billion yuan in total funding across all of 2025, while H1 2026 alone already exceeded 25 billion yuan — more than the full prior year in half the time.⚠️ Compiled estimate The reason money is piling in here is cost share — Morgan Stanley has estimated dexterous hands account for about 17.3% of Tesla Optimus's bill-of-materials cost, and industry participants generally put hardware cost at 10%–20% of the full unit, with high-end configurations potentially exceeding 20%. But capital enthusiasm and industry chaos coexist: shipment volumes across companies range from hundreds to 100,000 units, unit prices from just over 3,000 yuan to several hundred thousand yuan, and degrees of freedom from 6 to 42, all measured inconsistently. Xuanxiang Robotics CEO Su Jiuyu describes the current landscape as "unfocused" — there's no single dominant mainstream technical direction, application scenarios remain concentrated in research, and this year "could be called the first year of dexterous-hand mass production, and also the early stage of a bubble." One telling mismatch: the company with the highest shipment volume isn't necessarily the one with the highest revenue or valuation.Source: 36氪 source
Tianjin sets a target: over 20 billion yuan core output value for intelligent robotics by end-2028, Beijing-Tianjin-Hebei local-sourcing rate to 60% · industrial
Tianjin's Bureau of Industry and Information Technology, Development and Reform Commission, and Science and Technology Bureau jointly issued the "Tianjin Intelligent Robotics Industry Innovation Development Action Plan (2026–2028)" on August 10. The target for end-2028: core output value exceeding 20 billion yuan, average annual output-value growth over 15%, and the Beijing-Tianjin-Hebei regional local-sourcing rate raised to over 60%, alongside cultivating at least 3 companies with over 1 billion yuan in output value and 20 nationally recognized "little giant" specialized-and-innovative enterprises. Named technical priorities include integrated drive-control controllers, highly integrated joints, high-performance servo motors, high-precision reducers, multimodal fusion sensors, and dexterous hands, with an explicit goal of cultivating single-item manufacturing champions in precision transmission. On the application side, the plan calls for at least 30 demonstration application scenarios in public services. The plan's confidence is backed by existing figures — in H1 2026, Tianjin's above-scale intelligent-robotics enterprises posted total industrial output value of 6.849 billion yuan, up 15.2% year-on-year, with value-added growth up 18% year-on-year, an acceleration of 2.5 percentage points over Q1.Source: 证券时报 source
DYNA-2 claims pretraining on 1 million hours of human video, with zero robot data throughout · world-model
DYNA Robotics has released DYNA-2, architecturally a video-generation-based World-Action Model that, during pretraining on human video corpora, jointly optimizes for "predict the next frame" and "predict the next action." The company says the resulting contact-physics and spatial-reasoning capability transfers across embodiments — to fixed robot arms, humanoid prototypes, and five-fingered dexterous hands — despite none of these robots appearing in pretraining. The company's example: 13 minutes of data taught a pair of five-fingered hands to unscrew a bottle cap; it also claims the video co-training algorithm improves instruction-following scores by 133%.⚠️ Vendor-reported figures These are all self-reported internal comparisons (against the company's own prior VLA model, DYNA-1), with no third-party replication yet; the real point of interest is the claimed scaling law — if the human-video curve holds, scaling to 10 million hours would mean simply collecting more video rather than building out teleoperation robot fleets.Source: Unite.AI source
South Korean media report Samsung is mobilizing company-wide resources for its own humanoid, alongside world-model and actuator work · humanoid
ZDNet Korea reports that, beyond its subsidiary Rainbow Robotics, Samsung Electronics is independently developing its own humanoid robot, with R&D reportedly having reached "a fairly substantial level."⚠️ Report based on unconfirmed sources Two organizational moves are independently verifiable: last month's formation of RX (Robotics eXperience), a business promotion office reporting directly to Representative Director Roh Tae-moon; and the appointment of Vice President Lee Dong-gun from the Planning Group as head of the Robotics Strategy Group — he previously led robotics strategy at Hyundai Motor and was responsible for directing Boston Dynamics' operations. The report also says Samsung is developing the actuators and world model needed for its own humanoid, and plans to build a data factory at its Gumi site to feed existing production-line process data into robot training.Source: IT之家 source
UBTECH proposes an H-share incentive plan while applying for up to 2.6 billion yuan in bank credit · humanoid
UBTECH has proposed adopting an H-share incentive plan with a 10-year term and an authorization cap of 10% of total issued shares, and is applying for up to 2.6 billion yuan in RMB bank credit. This financing-and-incentive package lands the same day SAG listed UBTECH's H1 shipments at around 700 units and expressed reservations about its emotional-companion/bionic-styling approach — it reads more like groundwork for the company's next phase of cash burn.Source: 观点网 source
Automakers and full-machine manufacturers rush to set up new robotics entities: BAIC forms Yuanqi Physical Intelligence, UBTECH invests in a new company · embodied
BAIC Group has formed Yuanqi Physical Intelligence Technology Co., Ltd., with robotics R&D in its registered business scope; the same day, UBTECH and other investors formed a new robotics company with 10 million yuan in registered capital. Combined with GAC-linked Huilun Technology, which also completed a funding round today, spinning off embodied-AI business from the parent automaker into an independent entity has become standard practice — an independent entity makes market-based fundraising and equity incentives easier, while directly transferring the automotive supply chain's cost capabilities.Source: Gasgoo source
Waymo co-CEO: safety gains from a vision-only approach will plateau quickly · autonomy
Waymo's co-CEO publicly stated that safety improvements from a vision-only autonomous-driving approach will hit a ceiling quickly. This is a reaffirmation of the multi-sensor-fusion camp's position amid intensifying robotaxi scaling competition, an executive statement rather than new evidence. The same day, another report noted that while Waymo's robotaxi safety is improving, scaling up is exposing a new class of problems.Source: 디지털투데이 source
Tesla shows off its first Cybercab with an integrated Starlink antenna · autonomy
Tesla has unveiled its first Cybercab with a Starlink antenna integrated into the roof, with Musk saying the configuration will roll out to every vehicle. For a steering-wheel-free robotaxi, the point of an onboard satellite link is remote takeover and dispatch redundancy, not in-cabin entertainment — precisely the reliability issue regulators have been pressing on lately.Source: Electrek source
Hardware · Supply Chain
· Lens Technology · joint-module production line: Jiang Nan, President of Lens Technology's China operations and board secretary, disclosed that the world's first fully automated robot joint-module production line has been completed at its subsidiary Lens Intelligent (Changsha); the company has already delivered core components — humanoid and quadruped full units, joint modules, dexterous hands — to multiple leading robotics customers in China and overseas, with 2026 full-unit robot shipments surpassing 10,000 units, and has developed lightweight solutions including aluminum-magnesium alloy structural components. "World's first" per company statement source
· SGMW Industrial · magnesium-alloy structural components: delivered its first batch of magnesium-alloy "robot skeleton" components, migrating automotive lightweighting materials and processes onto humanoid robot bodies — pointing, like Lens Technology's aluminum-magnesium-alloy approach, at the same bottleneck: body weight reduction directly determines joint torque and battery-life budgets source
· Fullhan Microelectronics · robotics chips: the company confirmed in response to inquiries that it is already supplying chips to humanoid and industrial robotics companies, another Chinese chipmaker confirming shipments to robotics customers, following GigaDevice's robotics MCU disclosure source
Top comments (0)