Today's Highlights
· Former Qwen technical lead Lin Junyang starts a company, post-money valuation ~$2 billion
· Schaeffler completes validation of formed harmonic reducers, key-component cost down over 25%
· Tianjin's RD3 Ultra industrial humanoid rolls off the line, with existing production capacity in the hundreds
· 7 national standards for humanoid robot test methods launched for drafting
· Customs: industrial robot exports up 13.2% in the first 7 months, bionic robot exports up over 5x
· World model DreamX-Phi 1.0 takes first place in WorldArena 2.0 Track 1
I. Paper Progress
DreamX-Phi 1.0: action-conditioned video world model, tops WorldArena 2.0 Track 1 · world-model
Today's highest-trending paper in the community (HF↑77) targets an old problem with video world models: generated results look photorealistic but aren't faithful — a generation may look real while moving the wrong arm or losing track of the manipulated object. The approach injects each arm's SE(3) transform via PRoPE-style geometric encoding into attention to preserve arm identity and rigid-body motion structure, adds a lightweight depth branch to constrain scene geometry, and uses SAM3 masks paired with a frozen V-JEPA teacher to maintain object consistency throughout a grasp. The authors report the model ranked first on WorldArena 2.0 Challenge Track 1 and second on Track 2; the model and code will be open-sourced.
DreamX Team, Rui Chen et al. · arXiv 2608.13489 source
H2R-Bench: using human first-person videos to "manufacture" robot data — eleven models collectively fail · benchmark
"Replacing costly real-robot data collection with vast amounts of human first-person video" is a popular current approach to embodied data — this paper gives it its first systematic checkup: models rewrite human-hand demonstration videos into robot manipulation videos for a specified embodiment, scored across five dimensions — goal-state completion, action-event completion, functional-contact transfer, embodiment correctness, and image quality. Evaluating 11 leading video generation models across six manipulation task categories and two robot embodiments, the conclusion is that even top-performing models frequently fail on embodiment consistency and functional interaction — the cross-embodiment-transfer shortcut isn't viable yet.
Dingyi Rong et al. · arXiv 2608.13049 source · HF↑6
Temporal GRPO: fixing the "one wrong step, whole trajectory penalized" credit mismatch in VLA reinforcement learning · vla
Current GRPO-style VLA post-training spreads a trajectory's advantage value evenly across every action, so a rollout that gets every early stage right but fails at the end ends up penalizing the correct early actions too — what the authors call trajectory-level credit aliasing. The fix constructs detectable task stages, aligns each rollout to a stage interval, and only compares rollouts that enter the same stage against each other. On RoboTwin 2.0, both success rate and sample efficiency improve simultaneously; a controlled update on LIBERO-Long shows the improvement is concentrated at the stage where outcomes start to diverge.
Yao Zhou et al. · arXiv 2608.13026 source
VLAs hide a "progress bar" internally — a single linear probe can read it out · vla
VLAs are being deployed as general-purpose manipulation policies, yet the field has almost no tools to know what they represent internally or whether they're behaving normally at runtime. Using mechanistic-interpretability methods to probe the residual stream of π₀.₅, the authors find that "task progress" (normalized time remaining in the trajectory) is linearly readable, and that this signal already exists in the PaliGemma pretrained backbone before it ever touches robot data. A single linear probe generalizes to unseen tasks, though it isn't sufficient to steer the policy in return. Used as a label-free OOD detector to catch stuck tasks, it performs on par with SOTA methods — a lightweight path toward online monitoring.
Atiksh Bhardwaj et al. (Cornell) · arXiv 2608.13474 source
UniTexture: a single textured 3D object drops VLA success rate from 90% to 48% · vla
Prior adversarial attacks on robot policies were largely "one optimization per task"; this paper makes the attack generalize across tasks: gradients from the policy's action output are backpropagated through a differentiable renderer to the object surface's texture parameters, jointly optimizing a single shared texture across the distribution of tasks, instructions, states, and viewpoints. On OpenVLA and π₀.₅, average task success rate drops from a normal-condition 90.0% to 48.4%, and the texture transfers across tasks and models without re-optimization. As VLAs increasingly control physical bodies, this kind of general vulnerability — "place a single prop to induce unsafe actions" — deserves attention early.
Yukun Dai et al. · arXiv 2608.13453 source
GeoUP: porting a reconstruction model's geometric backbone into autonomous driving — SOTA across detection, occupancy, and depth · perception
Vision-only autonomous-driving perception commonly uses a backbone pretrained for semantic recognition, then patches in 3D geometry via downstream modules, making it hard for the shared representation to preserve metric scale and consistent scene structure. GeoUP adapts VGGT's reconstruction-oriented latent space to calibrated, streaming multi-camera scenes, splits cross-image interaction into self-attention, temporal, and view-wise pathways, and injects calibration-aware raymap encoding to supply metric scale; the same latent representation is decoded separately into depth, 3D detection, and semantic occupancy. It achieves SOTA on all three tasks across five datasets — nuScenes, Argoverse 2, Waymo, KITTI, and DDAD.
Longfei Xu et al. · arXiv 2608.13147 source
S2-HWM: 98.7% success rate on long-horizon surgical robot manipulation · world-model
Long-horizon surgical manipulation has extremely sparse rewards, and meaningful interaction changes happen at irregular points in time, yet existing world-model agents mostly "imagine" at primitive-step resolution, implicitly flattening out task progress. The authors have the model learn sparse event evidence from primitive-level latent trajectories, using an event-level manager to coordinate a primitive-level executor. This reaches a 98.7% success rate, 22.7 percentage points above a flat GAS DreamerV3 baseline.
Shuzhe Zhang et al. · arXiv 2608.13103 source
HumanoidVLN: filling the physics-simulation benchmark gap for humanoid vision-language navigation · benchmark
Existing VLN benchmarks mostly assume wheeled agents, sidestepping the physical constraints of bipedal walking, morphological differences across humanoid embodiments, and the first-person camera shake that walking itself causes. This physics-simulated benchmark covers multiple humanoid embodiments and is already compatible with leading methods including NaVILA, DualVLN, StreamVLN, and JanusVLN, offering a unified way to test "does it still walk with a different embodiment."
Quan-Dung Pham et al. · arXiv 2608.12860 source
Other papers today: FIRE-VLA converts unsolved failure cases from autonomous-driving VLA into privileged supervision, easing the no-signal problem GRPO faces when an entire group performs poorly (arXiv 2608.13395 source); Deliberate Practice computes a provably optimal allocation for skill learning under a limited practice budget (arXiv 2608.13415 source); Seeker learns region-of-interest attention purely from action supervision, with no gaze-point or object-category labels needed (arXiv 2608.13422 source); ContactGuard uses an action-conditioned latent world model to predict failure and abort before contact occurs (arXiv 2608.13438 source); EgoPHI is the first to jointly estimate dense contact maps and 3D force distribution on hand and object meshes from monocular RGB alone (arXiv 2608.13014 source); HumanTracker trains a human-preference-aligned humanoid motion-tracking metric from 12K motion pairs, exposing foot-slip and contact-timing errors that kinematic metrics miss (arXiv 2608.13555 source); WMRL replaces real-environment execution with a world model, speeding up reinforcement-learning training for autonomous research agents by 3-4x (arXiv 2608.12564 source); RoboSynChallenge is a competition-format unified evaluation of the generalization of manipulation policies trained on synthetic data (arXiv 2608.12416 source); FUSE lets agents actively decide "where to look" before judging affordance when functional cues are occluded (arXiv 2608.12683 source).
Open Source · Tools · Benchmarks
· Tier IV: the Japanese autonomous-driving company plans to develop its own automotive AI chip in-house and open-source the chip design, continuing its Autoware open-source approach source
· Liquid AI LFM2.5-VL-3B: a 3B-parameter on-device vision-language model built for reading screens, object localization, and on-device tool calling source
II. Funding & Deals
Pragmatik Labs (语用科技) | Angel round | ~$220 million | ~$2 billion post-money · world-model ⚠️ Reported, unconfirmed
Lin Junyang, former overall technical lead of Qwen (Alibaba's large-model line), announced his startup on X on August 12. The publicly reported investor lineup is co-led by Banyan Capital (Chinese VC firm) and HongShan (formerly Sequoia China), with backing from Tencent and the Shanghai Future Industry Fund; the funding size and valuation are drawn from public sources such as Tianyancha and people familiar with the matter, not company disclosure. The company positions itself around "world models and embodied intelligence," with two product lines — digital agents and physical agents — aimed at providing the underlying capability for next-generation agents that span both digital and physical worlds. Lin took over Qwen in late 2022, was promoted to Alibaba's P10 level last year, and left in March this year; as of the announcement the company has no commercial product or revenue, so the roughly 13.5 billion RMB-scale valuation rests almost entirely on the team and technical judgment. This is another instance — following ByteDance's Kong Tao and several core members of Tencent Robotics X — of top large-model talent migrating toward "agent + embodied."Source: Robot Insight (机器人大讲堂) source
INFIFORCE (原力无限) | Series A + A+ | ~RMB 1 billion · world-model
Led by Dunhong Capital (Chinese asset manager) and a top-tier state-owned investment platform, with follow-on from industrial investors including Zhejiang University's tech-transfer investment arm, Yandu State-owned Capital Holding, and Lishui municipal state capital, plus continued investment from existing shareholder CCV (Chinese venture firm). Funds go toward three areas: the embodied "brain" AtomBrain and causal world-model R&D, upgrading the full-stack AI infrastructure "DataGrid," and scaling delivery validation of multi-form robot bodies in real industrial settings. The company has bet on an ego-centric data approach since 2025 — using real human first-person operation to synchronously record vision, motion trajectories, spatial relationships, and environmental feedback, replacing costly real-robot data collection, to address the problem of models becoming unstable outside demo environments — and says it will launch its first embodied model built on high-quality ego data.Source: Gasgoo source
Light Origins (亮源新创) | Pre-A round | Hundreds of millions RMB · embodied
Led by Guoke Investment, with follow-on from China Merchants Capital and Xianghe Capital. This Beijing-based company focuses on foundation models for embodied intelligence, with funds going toward large-scale embodied model training, multimodal data infrastructure, full-stack software/hardware platform R&D, and team expansion. According to reports from ijiwei and iThome (Chinese tech-industry outlets), its founder is a former OpenAI algorithm specialist.Source: ijiwei source
Unitree Robotics | STAR Market IPO offline placement | Insurers allocated nearly RMB 1.03 billion · humanoid
Following the IPO price being set at RMB 150.80/share and online subscription oversubscribed by 8,288x, preliminary offline placement results show insurers as the dominant force in this subscription: 37 insurance companies and insurance asset managers were allocated about 6.8297 million shares, worth nearly RMB 1.03 billion, or 30.15% of the offline placement quantity; among the 2,617 valid quoting placement accounts under their management, various pension/annuity plan portfolios accounted for 1,635. The single largest recipient was Taikang Asset Management, whose 642 placement accounts together received 1.82 million shares, worth about RMB 274 million. The company's revenue for 2023–2025 was RMB 159 million, RMB 393 million, and RMB 1.699 billion respectively, with non-GAAP net profit attributable to parent of -RMB 18 million, RMB 78 million, and RMB 591 million. The listing date is still pending announcement from the Shanghai Stock Exchange.Source: 21jingji source
CablebotX (线控机器人, Nanjing Xiankong Technology) | Series A | Tens of millions RMB · industrial
Uses flexible cable-driven actuation to open up a new "spatial embodiment" category, targeting large-span, large-volume heavy-industry automation scenarios — areas long left unaddressed by traditional robotic arms, which are limited by reach and rigid structure.Source: DoNews source
Hengxuan Technology (恒璇科技) | Pre-A round | Tens of millions RMB · hardware
Led by Ninebot Capital and Daoyi Capital. Ninebot, itself a maker of full vehicles and balance/self-balancing scooters, moving into components investment is another example this round of industrial capital positioning upstream along the embodied-AI supply chain.Source: Sohu source
III. Commercialization & Deployment
Shiyun Technology's RD3 Ultra rolls off the line in Tianjin, industrial full-size humanoid enters local mass production · humanoid
The RD3 Ultra industrial-grade, full-function, full-size humanoid robot officially rolled off the line at the Tianjin Aerospace Digital Industrial Park, standing 174 cm tall with a standard battery life of 8 hours and a 10-second rapid battery swap, aimed at hazardous, repetitive, high-intensity specialized work. Alongside the production line sits a full-chain test center covering whole-machine calibration, walking-endurance aging, pre-shipment testing, and voice calibration: the walking-aging step requires the machine to walk continuously for over 1 hour, the line simulates composite terrain with a 5-degree slope (to be upgraded to 15 degrees), and replicates conditions like steps and walking backward up/down slopes. The company says it covers six application categories — station duty, industrial maintenance, specialized inspection, R&D training, environmental surveying, and emergency coordination — and is co-developing a health/elder-care humanoid solution with JD Robotics, with Shiyun providing the hardware body and motion control and JD providing the large model. Current production is in the hundreds of units, with the company planning to scale to a thousand-unit capacity next year. Compared with the recent wave of showroom and competition exposure, this kind of rollout — one that builds out a full testing line alongside it — comes closer to the threshold of industrial delivery.Source: Tianjin Daily source
SANY ships first batch of pure-electric unmanned mining trucks to South America, marking China's first entry into the Latin American unmanned mining-transport market · industrial
On August 12, SANY Heavy Equipment held a shipping ceremony at its Shenyang industrial park, with the first batch of SKT110Ei pure-electric unmanned mining trucks rolling off the line awaiting shipment; the delivery covers unmanned mining trucks, a supporting dispatch system, and maintenance services. The customer is moving toward automation under pressure from a long-term shortage of skilled mining-truck drivers and rising labor costs, and SANY says this is its first unmanned mining-truck project deployed in South America. As of July 2026, SANY Smart Mining says it has deployed over 300 unmanned mining trucks, with scaled commercial operations covering major open-pit mines in Xinjiang, Inner Mongolia, and Qinghai. China's unmanned mining-transport market is currently still dominated by third-party solution providers such as Trunk Technology (踏歌智行), Wisdom Motion (慧拓), and Easymile Robotics (易控智驾, unrelated to the French company EasyMile); OEMs going overseas with their own full-stack solutions is a separate path.Source: International Mining source
Kodiak receives California's first heavy-truck autonomous-driving testing permit · autonomy
The California DMV issued the permit on August 13, the first issued since the state passed its heavy-duty autonomous vehicle testing regulations on April 28, allowing Kodiak to test autonomous trucks with a safety operator on board. California's rules follow a phased path: testing with a safety operator → driverless testing → driverless deployment, and this is only the first phase. Founder and CEO Don Burnette said this lets the company "begin the first phase of expanding autonomous freight coast to coast." California is a major U.S. freight hub that has long kept heavy trucks outside its autonomous-vehicle regulations, so this permit's symbolic significance outweighs the current fleet size.Source: Truck News source
China Post's Guangzhou sorting center lifts embodied-robot parcel-feeding rate to 1,200 pieces/hour · embodied
The Guangzhou postal district center's Jianggao processing workshop handles an average of 6.5 million pieces of mail daily, with peak-day volume exceeding 10 million; in the past, peak periods relied mainly on manual sorting. A batch of embodied robots began installation this March, and after more than five months of operation and upgrades, parcel-feeding efficiency rose from 300 to a peak of 1,200 pieces per hour, reaching over 60% of manual processing efficiency. The robots use 3D vision to identify and locate parcels, paired with flexible dexterous hands adapted to common packaging materials. "60% of manual" may not sound high, but in a logistics-sorting role that runs 24 hours a day with fluctuating labor availability, what it measures is whether the robots can hold the position on a sustained basis, not peak-moment performance.Source: China Youth Online source
Cadillac's new XT5 PHEV becomes the first mass-production model equipped with the Momenta R7 world model · autonomy
SAIC-GM Cadillac announced on August 12 that the new XT5 PHEV is the first mass-production luxury hybrid SUV equipped with the Momenta R7 world model, with its driver-assistance capability usable from delivery; showroom units have started arriving at dealerships. The vehicle is equipped with a 1.5T engine (max power 115kW) and dual front/rear motors (160kW/110kW), paired with a 35.5kWh Ultium ternary lithium battery, giving a WLTC pure-electric range of 155 km. This production milestone traces back to July 8 — the day Momenta listed on the Hong Kong Stock Exchange main board, when Cadillac announced it would debut the industry's first mass-production application of the R7 world model. World models moving from papers and demos into a mass-production car you can actually buy is one of the few places this has landed so far.Source: Gasgoo source
AutoStore signs global strategic supply agreement with Amazon, but with no purchase commitment · industrial
Norwegian warehouse-automation company AutoStore announced on August 13 a global strategic supply agreement with Amazon, establishing framework terms for Amazon's worldwide procurement of its automation systems. The company emphasized that the agreement currently includes no purchase commitment, with amount, potential order volume, deployment locations, and timeline all undisclosed. The announcement was classified as inside information under EU market abuse regulations. AutoStore's system has robots walking atop dense grids of aluminum storage bins, retrieving bins and delivering them to workstations; the company says it now has roughly 2,000 systems installed across 68 countries. The agreement itself doesn't confirm a large order, but by settling the underlying terms once instead of negotiating deal by deal, it clears a channel for potentially larger-scale deployment later.Source: Robotics & Automation News source
IV. Industry Developments
Pony.ai expands partnership with Uber, plans to deploy over 2,000 robotaxis across five European cities · autonomy ⚠️ Plan stage
Pony.ai and Uber announced an expanded partnership, planning to deploy over 2,000 robotaxis across five European cities and extend into the Middle East, with specific timelines and cities not yet announced. The two companies first partnered in May 2025 focused on the Middle East, then shifted to Europe, launching commercial service in Zagreb, Croatia with local partner Verne; Uber has now allied with over 30 autonomous-driving companies.Source: TechCrunch source
LG signs MOU with Nvidia, first bipedal humanoid set to debut Q1 next year · humanoid ⚠️ Plan stage
LG Group Chairman Koo Kwang-mo and Nvidia CEO Jensen Huang signed a strategic-cooperation memorandum on August 13, covering three areas — humanoid robots, AI factories, and mobility — and showed a model of the first bipedal humanoid robot, planned to debut in the first quarter of next year. The unit is being developed on Nvidia's Isaac GR00T platform with a Jetson Thor compute module; LG is integrating components from its own subsidiaries including LG Electronics, LG Energy Solution, LG Innotek, and LG Display.Source: Chosunbiz source
Schaeffler completes validation of formed-process harmonic reducers, key-component cost down over 25%, mass production in 2027 · hardware
Harmonic reducers are core to humanoid robot joints, and actuators account for roughly half of a robot's total manufacturing cost. The current mainstream manufacturing method, precision machining, is time-consuming and capital-intensive, and is seen as the main bottleneck to scaling humanoid production over the next several years. Schaeffler has transplanted a forming process refined over years in the automotive sector, replacing material removal with high-pressure single-step forming; manufacturing time for key components drops from minutes to seconds, manufacturing cost falls by over 25%, and material consumption drops by over 75%. The company says it has completed validation testing of the formed harmonic reducer, laying the groundwork for mass production in 2027, starting in Germany before expanding to other regions; over the past decade it has already supplied more than 2 million formed harmonic reducers to major automotive markets worldwide. The cost-reduction narrative for humanoid robots has so far mostly stayed at the level of headline unit prices — once the reducer, the single largest cost item, has its manufacturing paradigm replaced, that's the sustainable part of the story.Source: Robotics & Automation News source
Customs: industrial robot exports up 13.2% in first 7 months, smart bionic robot exports up over 5x in half a year · adjacent
Wang Jun, deputy head of China's General Administration of Customs, disclosed that China's industrial robot exports grew 13.2% year-on-year in the first 7 months of this year; since a dedicated customs tariff code was introduced for smart bionic robots this January, the category's export value has grown over 5x within half a year. Over the same period, China's total goods trade reached RMB 30.13 trillion, up 17.3% year-on-year, with electromechanical product exports at RMB 11.12 trillion, up 21.2%. Data from the Ministry of Industry and Information Technology shows Chinese-developed quadruped robots hold nearly 70% of global sales share, and Chinese humanoid robot products number over 400 models, more than half the global total. At the local level, Beijing's service-robot output grew 2.3x and industrial-robot output grew 75.5% in the first half of 2026; key humanoid-robot manufacturers in the city produced nearly 6,000 units, up nearly 10x year-on-year. The dedicated tariff code itself is worth noting as a signal — only once a statistical category exists does this segment get trackable export data for the first time.Source: Sina Finance source
7 national standards for "Humanoid Robot Testing Methods" launched for drafting · adjacent
On August 13, the launch meeting for the "Humanoid Robot Testing Methods" series of national standards was held at Wuhan's Optics Valley, with 7 standards — including general principles, environmental perception, decision-making and planning, and motion control — launched for drafting simultaneously. Over 200 representatives from companies including Unitree Robotics, Xiaomi Robotics, MagicLab (Chinese humanoid-robot maker), ZTE, DEEP Robotics, Horizon Robotics, Linker Bionics (Chinese dexterous-hand maker), and Galbot (Chinese embodied-AI startup), along with universities including Wuhan University and Huazhong University of Science and Technology, attended. The Hubei Provincial Electronic Information Standardization Technical Committee's humanoid robot and embodied intelligence standards working group was also inaugurated the same day, with an initial roster of 29 member organizations. Liu Xin, secretary-general of the Humanoid Robot Sub-Technical Committee under the National Robotics Standardization Technical Committee, said testing methods specify "how to test and how to determine whether it truly meets the standard" — equivalent to a nationally-recognized exam syllabus; currently, companies' technical standards and test methods differ from one another, making scaled replication difficult. At a time when whole-machine specifications are inconsistent and success rates and battery life are each reported differently, unifying test methods is closer to the industry's real problem than yet another new machine.Source: Optical Communication Pro (光通信Pro) source
Hyundai Motor Group signs agreement with South Korea's Army and Ministry of Trade, Industry and Energy to advance defense physical AI · adjacent ⚠️ Plan stage
The three parties signed a memorandum on August 14 at the Army headquarters in Gyeryongdae, South Chungcheong Province, to evaluate applying robotics and physical AI to defense logistics. The first scenarios slated for evaluation include logistics automation, supply transport, and perimeter/security patrol; selected projects will be validated for reliability and usability at a testbed built within the Army, and joint civil-military research will take place at the "AX Hub" in Pangyo, set to be fully operational this year. The cooperation also covers hydrogen mobility and infrastructure, along with ecosystem links to the "Saemangeum AI Valley" robotics manufacturing cluster and AI data centers. The driving factor is a shrinking pool of military personnel from demographic change — also the shared backdrop behind several Japanese and South Korean automakers' recent collective pivot toward robotics.Source: Chosunbiz source
Engine AI (众擎机器人) co-founds new company, business scope includes smart unmanned aerial vehicle manufacturing · humanoid
Shenzhen-based Engine AI (humanoid robot maker) co-founded a technology company with partners, with a business scope including smart unmanned aerial vehicle manufacturing. Humanoid-robot makers branching sideways into low-altitude aerial vehicles is a recent side trend — what's shared is the underlying capability of motors, joint modules, and motion control, rather than building a new product category from scratch.Source: EEPW (电子工程专辑) source
Hardware · Supply Chain
· Micro planetary roller screws: Nuoshi Robotics (Chinese precision-actuator maker) has mass-produced planetary roller screws just 1.5mm in diameter with a 5.5mm nut diameter, C5-grade precision, ±0.01mm repeat positioning, and 50N max load; the integrated micro linear actuator outputs 15kg of thrust within a USB-drive-sized footprint. Its Suzhou plant, running 24 hours at full capacity for an annual output of 3 million units, still can't keep up with demand; the company aims to raise annual output to 10 million units by year-end. The founder says demand has extended from dexterous hands to wrists, forearms, and even neck and face, with neck degrees-of-freedom requirements rising from one or two to four to six source
· Bionic e-skin: a Chinese research team has developed bionic electronic skin that mimics human tactile perception; according to the researchers, current application directions include tactile sensing modules for humanoid dexterous hands, tactile data-collection gloves, and smart cockpits source
Top comments (0)