Today's Highlights
· Enigma emerges from stealth with a $70 million seed round, opening 100+ of its own robots to be teleoperated by anyone online
· Caocao Mobility opens driver-seat-safety-driver-free Robotaxi testing in Hangzhou, where a fleet of hundreds of vehicles is already deployed locally
· Jack Technology, a leading sewing equipment maker, orders 2,000 humanoid robots in one go, partnering with Siemens to overhaul garment production lines
· Jianlan Luo's team embeds a world model into a VLA with τ0-VLA, completing a 25-step room-tidying task in one run
· Surgical robot exports surged 3.3x to 480 million yuan in H1, with robotics named among China's "new new three" export categories
I. Research Progress
τ0-VLA: Embedding a World Model into the High-Level Layer of a VLA, So Robots "Imagine Before Acting" · vla
Long-horizon tasks have long been stuck on accumulated error and incomplete memory — repeatedly picking up the same piece of clothing, skipping the step of "opening the laundry basket lid," or drifting from the original instruction partway through. τ0-VLA's approach is to embed a world model in the slow-thinking layer of a dual-system architecture: a proposal model generates multiple candidate subtasks, a world model imagines what the scene will look like after each subtask executes, a value model scores the imagined futures and searches backward recursively, and a reflection model decides which candidate to finally commit to — the key design is that the world model only predicts and never scores, since otherwise the model would be grading its own work. This moves the kind of test-time compute common in large language models into embodied AI, and it happens only at the subtask decision layer rather than the low-level action layer. The paper sets up seven real-robot tasks and, on the Agibot G1, completes a 25-step (about 8-minute) room-tidying task, a 22-step tomato-and-egg stir-fry, and 14-step meal prep.
Jianlan Luo's team (Shanghai Innovation Institute × Agibot Embodied AI Center) · Analysis: 具身纪元 (Embodied Era, a Chinese embodied-AI newsletter) source (WeChat, CN)
CD-LAM: World Model Video Looks Smooth, But Often Ignores the Action Commands You Give It · world-model
Latent Action Models (LAM) were originally meant to let world models pretrain on massive amounts of unlabeled human video, but a causal analysis by Aether AI and UC San Diego found a critical flaw: LAM encodes background, camera motion, and scene continuity all together into the "latent action," so the world model learns to shortcut by guessing the next frame from visual continuity — zero out the actions entirely, and the generated robot arm still moves. CD-LAM leaves the world model's main body untouched, purifying the latent action representation upstream instead, using three objectives — embodiment-centric reconstruction, action-centric contrastive learning, and latent-space calibration — which cuts action error by over 30%, and post-training needs only 3k-6k fine-tuning steps, about 10% of the baseline's compute cost. The team also proposes an FDCE metric, since PSNR barely correlates with action accuracy.
Yufan Wei et al. (Aether AI × UC San Diego) · arXiv 2607.09185 source · Analysis: 机器之心 (Synced, a Chinese AI-news outlet) source (WeChat, CN)
LDA-1B: A World Model That Puts Low-Quality Data to Work Too, Each Type Doing Its Own Job · world-model
High-quality expert trajectories are perpetually scarce, so LDA-1B's approach is to divide labor across heterogeneous data: action-unlabeled human video learns visual prediction, low-quality trajectories learn dynamics, and high-quality trajectories learn both policy and dynamics simultaneously. The accompanying EI-30k dataset is also released — roughly 30,000 hours of mixed real and simulated human-and-robot trajectories with action alignment. The architecture is a multimodal diffusion MoE Transformer, with action and vision experts fused at self-attention; borrowing from LeCun's JEPA idea, it doesn't predict pixels but instead supervises in DINOv3 latent space. In dexterous-hand manipulation experiments, it outperforms Pi0.5 and NVIDIA's GR00T.
Peking University × Galbot (Chinese embodied-AI robotics company) × Institute of Automation, Chinese Academy of Sciences × BAAI (Beijing Academy of Artificial Intelligence) × Tsinghua University × Sun Yat-sen University × NVIDIA · Analysis: 机器人解剖师 (Robot Anatomist, a Chinese robotics newsletter) source (WeChat, CN)
Pigey: No Retraining Needed — Just Add a "Foreman" Layer, and LIBERO-PRO Score Quadruples · vla
The mainstream approach today is to cram perception, planning, success detection, and recovery all into one policy via large-scale pretraining; this paper goes the opposite way, splitting the capability into a general language-conditioned control policy plus a high-level orchestrator. Pigey does closed-loop planning, breaks down subgoals, dispatches low-level instructions, verifies outcomes from observations, and recovers from failures — without any additional data collection or post-training, and it can drive off-the-shelf VLA policies as-is. On LIBERO-PRO, this raises performance from 12.8% to 53.3%, and on real robots it lifts a frozen policy on inference-constrained tasks from near zero to over 90%. The authors name the gap between executing a frozen motor skill standalone versus inside an agentic loop the "orchestration gap."
Liane Galanti, Dhruv Shah, Tri Dao · arXiv 2607.21725 source
Robot-Factored World Model: Stop Making the World Model Learn "How Instructions Become Actions" · world-model
An action's influence on future video actually happens in two steps — first the robot's own body and controller realize it as motion, then the scene responds through contact. Conditioning directly on the action command forces the world model to learn the first step too; conditioning on the recorded future state instead leaks the interaction outcome the model is supposed to predict. This paper moves two robot-specific factors outside the model: instructions first pass through the robot's own controller and kinematics to get a nominal trajectory, which is then rendered via URDF into explicit robot geometry, so the model perceives action only in the form of "visible robot geometry." Experiments show this interface outperforms vector-conditioning baselines, generalizes to embodiments unseen at inference time, and can even retarget human hand demonstrations, rendering them as robot geometry to generate manipulation videos.
Byungjun Kim, Taeksoo Kim, Hyunsoo Cha, Hanbyul Joo · arXiv 2607.22535 source
ViTacWorld: Using a World Model to Directly Generate "Vision + Touch" Trajectories · world-model
Contact-rich manipulation depends on physical cues invisible to a camera, but real tactile data collection is expensive, hardware-bound, and limited to a single scenario. ViTacWorld's key observation is that tactile signals are directly anchored to physical contact, so the sim2real gap is smaller than for vision alone — this makes it possible to scale up data by combining public real tactile data with self-built simulation environments, pretraining first and then fine-tuning via real-robot policy rollouts. Given an action, the model simultaneously predicts temporally aligned visual observations and tactile feedback — synthesizing rollouts to improve downstream tactile policies on one hand, and evaluating policies under controlled action sequences on the other. The authors call this the first framework to use a world model for vision-tactile-action trajectory generation and policy evaluation.
Yunao Huang et al. · arXiv 2607.22530 source
Xiaomi-Robotics-U0: Making the World Model a Producer of Training Experience, Not Just a Video Generator · world-model
The hard part of scaling embodied data was never the volume of video, but whether task relationships still hold once you change the background, lighting, or object instances — the same object across multiple cameras must correspond to the same 3D scene, and the relative position between arm and object can't drift frame by frame. U0 uses a multimodal autoregressive backbone to put ordinary image tasks, embodied scene generation, cross-scene transfer, and embodied video generation all into a single unified model, and during annotation it first separates out semantic fields such as workbench/object/lighting/background, then segments subtasks and keyframes based on end-effector pose and gripper signals. Its comparison against GPT-Image-2 focuses on whether depth constraints and multi-view correspondence hold up, rather than single-frame visual quality.
Xiaomi Robotics team · arXiv 2607.11643 source · Analysis: 具身智能论文日记 (Embodied AI Papers Diary, a Chinese newsletter) source (WeChat, CN)
ACE-Brain-0.5: An 8B Backbone That Fuses Spatial Perception, Planning, Manipulation, and Self-Monitoring into One Brain · vla
End-to-end VLAs are good at producing actions but weak on spatial reasoning and long-horizon planning, while LLM-based robot agents can orchestrate tools but fail to learn a unified robot representation. ACE-Brain-0.5 uses a hybrid Transformer to fold five cognitive functions into a single model, with a backbone of only 8B parameters, and introduces a dual-timescale control interface — an LLM decoder handles semantic planning at low frequency, while a lightweight fast-vision pathway injects real-time features into a flow-matching action expert at high frequency. To suppress cross-task interference from joint training on heterogeneous multi-task data, the team adds a "reactivation" stage on top of its original SSR strategy, upgrading it to SSR+. It reaches a 98.2% average success rate on LIBERO and 82.3% on SimplerEnv-Bridge.
Ziyang Gong et al. (ACE-Brain Team) · arXiv 2607.04426 source · Analysis: Green生态智能机器人 (Green Ecological Intelligent Robotics, a Chinese newsletter) source (WeChat, CN)
Other papers today: Ordered Action Tokens designs structured, ordered action tokens for visuomotor policies, avoiding the dilemma between overly long parsed discretization and unstructured implicit tokenizers (arXiv 2607.21670 source); ACME releases a cross-cultural, cross-embodiment social navigation dataset, filling a gap in regional and interaction diversity (arXiv 2607.21964 source); GRACE combines diffusion policies with MPPI posterior mean estimation, satisfying hard constraints like collision detection and joint limits without needing a differentiable guidance cost (arXiv 2607.21661 source); a survey on Progress Reward Modeling systematically reviews process reward design in robot learning (arXiv 2607.21655 source); SiPhy infers physical properties like mass, stiffness, and elasticity from a single image (arXiv 2607.22355 source); Mag4D-SLAM releases a multimodal 4D geomagnetic localization and mapping dataset with repeated traversals (arXiv 2607.21986 source); a skin-covered robotic arm's tactile "withdrawal reflex" pipeline underwent human-factors evaluation (arXiv 2607.22249 source).
Open Source · Tools · Benchmarks
· NVIDIA Cosmos-H-Dreams: a distilled surgical world model runs at roughly 160 fps on a single RTX PRO 6000 (versus roughly 10 fps for the teacher model); weights and serving code were released July 23. There's no physics engine underneath — it learns visual dynamics directly from surgical video and robot kinematics. The training set deliberately retains failures and out-of-distribution clips such as dropped needles, missed stitches, and loose knots — a simulator meant to score policies must be able to reproduce the consequences of bad actions. The currently released checkpoint only does tabletop suturing on the da Vinci Research Kit, at 288×512 resolution and a 10 Hz command frequency source
· Beijing humanoid open dataset: Beijing's Yizhuang district disclosed that its embodied AI open-source dataset has surpassed 10 million downloads globally source
II. Financing and Deals
Enigma (US/Israel) | Seed Round | $70 Million · embodied
Led by Index Ventures and Ribbit Capital, with participation from Conviction Partners' Sarah Guo (the company's own press release cites $71 million, while TechCrunch reported $70 million). Less than a year old, this research lab is taking a path opposite to its peers: while others race to build foundation model capability — scraping web video, running simulations, using sensor-equipped gloves to capture human motion — Enigma studies how humans interact with robots, betting that this will yield an intuitive interface and perhaps even a different kind of robot brain. Alongside its emergence from stealth, it's opening up 100+ of its own robots, housed in hangars in Israel and California, letting anyone in the world teleoperate them online — they can paint, spar with swords, and perform simple chemistry experiments with flasks; both the robotic arms and the underlying models are described as self-developed. The two co-founders, Jonathan Jacobi (Microsoft's youngest-ever employee) and Gal Niv, met at a teen hacking competition and served together in Israel's Unit 8200; neither had previously worked in robotics.Source: TechCrunch source
Zhigu Tianchu (Shenzhen) | Strategic Round | Nearly 100 Million Yuan · adjacent
Led by China Merchants Capital, followed by Leap Capital, with existing investors Yunqi Partners and Qifu Capital oversubscribing. Founded in 2018, this intelligent cooking robot company is a rare example among embodied AI firms in that it has revenue over 100 million yuan and is already profitable; it expects 2026 revenue to maintain 300% growth over 2025. Its products are already used in smart manufacturing, elder care, campus canteens, and fast-food chains, with a recent deployment in the logistics canteen of one of the world's largest new-energy-vehicle makers, and bulk rollout in school canteens in the Beijing area. Since late 2025, daily overseas client visits have roughly tripled and the deal-conversion rate exceeds 50%, with products already deployed in Europe, North America, Australia, Southeast Asia, and Japan/Korea. The company is about to rebrand as "Qianyi Intelligence."Source: 机器人前瞻 (Robot Foresight, a Chinese robotics newsletter) source (WeChat, CN)
Zhengqi Future | Three Rounds of Angel/Series Funding in 8 Months | Several Hundred Million Yuan · autonomy
Core investors include CDH Investments' VGC fund and Linear Capital, alongside industry investors SAIC's Hengxu Capital, Bǎiruì Capital, Zhengxuan Investment, and Chow Tai Fook Enterprises, with existing shareholders continuing to invest and the round oversubscribed. Founder Wu Shulin spent 15 years at Huawei and previously served as vice president of Baidu's Intelligent Driving Group, and he's targeting door-to-door short-haul mobility tasks — the "urban traffic capillaries" spanning from bedroom to elevator, corridor to neighborhood path. His thesis is that short-haul travel accounts for over 70% of total human travel, yet autonomous-driving penetration there is nearly zero, while hundreds of billions have already been invested in road-based autonomous driving covering only the structured 30%. The product line is the QUORRA-branded short-haul mobility robot, featuring a proprietary DoorMind world model and a FLAT locomotion platform with four independently steerable wheels; the first model, QUORRA X5, has already secured commercial orders from more than a dozen overseas countries and regions, and will open flagship stores on Tmall and JD.com this year.Source: 融中财经 (Rong Zhong Finance, a Chinese finance outlet) source (WeChat, CN)
Standard Robots (Wuxi) | Third Filing for Hong Kong IPO · industrial
CITIC Securities and Guotai Junan are joint sponsors. By 2025 sales volume, it ranks as China's fourth-largest provider of industrial intelligent mobile robot solutions, with 4.0% market share; as of July 23, its solutions had been adopted by over 400 clients globally, ranking second in the 3C electronics sector, third in automotive, and fifth in semiconductors. Financially, revenue grew from 162 million yuan in 2023 to 301 million yuan in 2025, but losses narrowed from 100 million yuan to 45.144 million yuan over that period before widening again to 202 million yuan; in the first four months of 2026, revenue was 107 million yuan with a loss of 61.835 million yuan.Source: 智通财经 (zhitongcaijing.com) source
Linghu Intelligence (Shenzhen) | New Round | Tens of Millions of Yuan · hardware
Sole investor Haiyuan Capital, with funds going toward R&D iteration of joint modules and multiple motor models plus capacity expansion. Founded in 2016, its products span joint modules, educational robots, industrial robots, and micro motors, with the motor lineup covering coreless, permanent-magnet synchronous, frameless torque, and brushless DC types across the board; its ultra-small-diameter coreless motors are already capable of mass production and are widely used in dexterous hands. It has already partnered with companies including UBTech.Source: 机器人前瞻 (Robot Foresight) source (WeChat, CN)
Axis Robotics (US) | New Round | $12 Million · adjacent
The funds will go toward building a compounding data engine for physical AI.Source: Yellow.com source
III. Commercialization and Deployment
Caocao Mobility Opens Driver-Seat-Safety-Driver-Free Robotaxi Testing, Fleet of Hundreds Already Deployed in Hangzhou · autonomy
Caocao Mobility announced on July 27 that it has officially opened driver-seat-safety-driver-free Robotaxi testing, conducted on real city roads in Hangzhou's Binjiang District, using a specific-stop-triggered dispatch mechanism to validate the stability of safety-driver-free operation under small-scale, controllable conditions. All test vehicles are connected to the "Caocao Zhixing RAS Remote Safety Service Platform," which handles full-process safety monitoring, emergency call response, remote real-time assistance, and anomaly handling. In April this year, the company became the first in Hangzhou approved to conduct safety-driver-free intelligent connected vehicle road tests, and it now has a Robotaxi fleet of hundreds of vehicles deployed in Hangzhou, supported by a "green smart transit island" integrating automatic battery swapping, automatic cleaning, and vehicle servicing. As the commercialization vehicle for Geely Holding's Robox strategy, its co-developed EvaCab — China's first natively-developed Robotaxi — has already been unveiled, with mass production to replace the current retrofitted vehicles planned for 2027. Set against this week's broader context, this follows the mass-production delivery of Dongfeng Nissan Venucia × Hello's HR1 and Hong Kong's first fully driverless license, marking another step in China's Robotaxi progression from "having a license" to "removing the safety driver."Source: Gasgoo source
Jack Technology Orders 2,000 Humanoid Robots in One Go, Partners with Siemens to Overhaul Garment Manufacturing · industrial
Chinese industrial sewing machine manufacturer Jack Technology is partnering with Siemens to integrate AI, digital engineering, and humanoid robots into garment production, aiming to boost manufacturing efficiency by at least 30% (the efficiency figure is the company's stated target). As part of this push, Jack Technology has ordered 2,000 humanoid robots customized for garment manufacturing, which the company describes as one of the first large-scale humanoid deployments in the industry — the supplier and delivery timeline for the order have not been disclosed. Garment manufacturing is a classic labor-intensive setting with small batches and tight deadlines, and humanoids are seen as attractive here because they can step directly into workstations designed for humans and use existing tools and equipment without major production-line changes. The same report also mentions a two-week pilot at Siemens' Erlangen electronics factory in Germany: a wheeled humanoid moved 60 bins per hour, could handle two box sizes, ran over 8 hours a day, and achieved an autonomous pick-and-place success rate above 90% — ⚠️ demo-level figures, as a two-week pilot's duration and scale don't yet constitute a mass-production benchmark.Source: Interesting Engineering source
Value Truck to Deploy Aurora's Second-Generation Driverless Trucks · autonomy
Following Aurora's US launch of its second-generation driverless truck on July 22, this announcement lands on the customer side — freight company Value Truck announced it will deploy this generation of trucks; fleet size and rollout timeline were not disclosed in the announcement.Source: Business Wire source
Momenta's Robovan Launches in Suzhou, Extending the Same World Model from Passenger Cars to Driverless Logistics Vans · autonomy
Momenta's Robovan is now running in Suzhou, marking the extension of its world model foundation platform from passenger cars and Robotaxis to Robovans. The company launched its R7 world model into mass production this past April, with a strategy of using a single large model to support passenger cars, Robovans, Robotrucks, and Robotaxis at scale — the opposite of most players' approach of building separate stacks per vehicle type.Source: Sina Finance source
BAUHAUS Brings in XYZ Robotics to Automate Its Entire Receiving Process · industrial
German building-materials retailer BAUHAUS has partnered with XYZ Robotics (Chinese robotic-grasping company) to fully automate its receiving process. Receiving has long been a corner of warehouse automation that companies avoided — incoming boxes come in varied shapes with no fixed sequence, demanding more general-purpose grasping ability than finished-goods picking.Source: Yahoo Finance source
Gatwick Airport to Launch the UK's First Robotic Valet Parking Service Next Month · adjacent
London's Gatwick Airport has partnered with French company Stanley Robotics to launch fully automated robotic valet parking at the South Terminal next month: after a driver parks in a designated space, a robotic platform slides under the vehicle, lifts its tires, and transports it to a secure parking area — the keys never leave the driver's hands. The airport says the system is compatible with up to 95% of vehicle models, with entry and exit controlled by license-plate recognition and round-the-clock monitoring. The real economics here aren't about the experience but the land — robots can park cars more densely than humans, raising capacity without expanding facilities. Stanley Robotics has already deployed this at Lyon Airport in France, with Basel-Mulhouse Airport set to follow.Source: Secret London source
IV. Industry Developments
Robotics Named Among China's "New New Three" Export Categories: Surgical Robot Exports Surge 3.3x, Market Expands from 23 to 49 Countries · adjacent
According to CCTV Finance, robotics, artificial intelligence, and innovative pharmaceuticals are emerging as China's "new new three" export categories. By customs figures, cleaning robots and intelligent bionic robots together exported 18.09 billion yuan worth of goods in H1; industrial robot exports reached 6.29 billion yuan, up 18.6%, sold to 141 countries and regions worldwide; surgical robot exports reached 480 million yuan, surging 3.3x, with the export market expanding from 23 to 49 countries — the fastest growth is precisely in the category with the highest unit price and the toughest certification barriers. Data from Qichacha shows that as of July 27, China had 1.1994 million existing robotics-related companies; 2025 saw registrations surpass 200,000 for the first time, reaching 277,700 for the year, up 39% year-on-year, with 155,500 already registered so far this year, up 19.24% year-on-year. Over half of these companies were founded within the past three years, with those aged 1-3 years making up the largest share at 29.33%. The flip side of these numbers is an industry structure that's extremely young, with consolidation not yet underway.Source: 观点网 (Guandian.com, a Chinese real-estate/business news site) source
Tesla Buys Its First Full-Body Virtuix Simulation System for Its Optimus Division · humanoid
Virtuix (NASDAQ: VTIX) announced that Tesla has procured its first Omni One Enterprise system for its Optimus humanoid robot division — this is the product's first enterprise order. The Omni One is an omnidirectional treadmill that lets users walk naturally through virtual environments within a compact footprint; the company's official use cases include training, simulation, and humanoid robot teleoperation. Virtuix's previous partners have mostly been in consumer, defense, and medical sectors, including Meta, NASA, and all four branches of the US military. The purchase amount was not disclosed, and Tesla has not specified its intended use.Source: Yahoo Finance source
Waymo's Urban Friction Bill: $9,325 in Parking Tickets in Austin, San Francisco's Mayor Calls for Tighter Rules · autonomy
Waymo's driverless vehicles in Austin have racked up $9,325 in parking fines, while a Waymo-involved accident in Los Angeles injured multiple people around the same time, and San Francisco's mayor is calling for stricter rules after traffic disruptions. These bills aren't large in isolation, but they represent the first quantifiable external costs as Robotaxis move from "can it drive" to "can it permanently occupy public urban space."Source: Sina Finance source, InsideEVs source
Boston Dynamics Embeds Google Cloud's Gemini into Spot · embodied
Boston Dynamics has integrated Google Cloud's Gemini into its quadruped robot Spot. Spot's main battleground is industrial inspection — this week it was deployed for autonomous inspection at the Copper One copper mine — and adding a multimodal large model points toward turning "walk a preset route" into "understand an on-site instruction and then decide what to look at."Source: MarketScale source
Zhiyuan Robotics (Chinese humanoid startup) Continues Hong Kong IPO Push: Revenue Grew from 300,000 Yuan to 1.05 Billion Yuan in Three Years, Five Business Lines Each Spun Off with Their Own Funding · humanoid
Following its formal Hong Kong IPO launch on July 24, this update covers its financial and organizational foundation (already reported). Chairman Deng Taihua revealed in April this year that since the company's founding three years ago, revenue grew from 300,000 yuan in its first year to over 60 million yuan in its second year, and surpassed 1.05 billion yuan in 2025. Organizationally, starting in April 2025 Zhiyuan spun off its five business segments — dexterous hands, data collection, leasing services, quadruped robots, and commercial cleaning — into independent subsidiaries, each of which has completed its own independent financing round.⚠️ Single-source valuation figures (roughly $20 billion pushing forward the IPO, with a cornerstone target of HK$40-50 billion) remain unconfirmed by the company.Source: 东方财富 (Eastmoney) source
BYD Names Its Humanoid Robot "Xiaodi," Debuting in Zhengzhou in Early August · humanoid
Following last week's announcement that it was entering the humanoid space, BYD has revealed the name and specifications of its first humanoid robot: 1.61 meters tall, weighing 58.5 kg, with 31 degrees of freedom, capable of fine-grained grasping with its hands, and equipped with a multimodal perception system supporting multiple languages and Chinese dialects. It's positioned as a showroom sales assistant, handling tasks like greeting customers and explaining vehicle models, with its first public appearance planned for early August in Zhengzhou. Overseas reports say each dealership is planned to be equipped with two units.Source: 猫头鹰人形机器人 (Owl Humanoid Robotics, a Chinese newsletter) source (WeChat, CN), Autohome source
Unitree's Wang Xingxing: Robots Entering the Home Still 5 to 10 Years Away · humanoid
In an interview, Wang Xingxing said it will take 5 to 10 years for robots to enter households the way ordinary appliances have, with safety and ethical issues being what will need the most gradual adaptation. ⚠️ Stated opinion — worth comparing against his earlier remark this month that embodied AI's "ChatGPT moment" could arrive in as little as two to three years; the two statements refer to different things — a technological inflection point versus household adoption.Source: China News Service source
Hardware · Supply Chain
· Harmonic reducers: PwC's "2026 Smart Robotics Industry Development White Paper" states that domestic (China-made) localization has already exceeded 75%, driving industry costs down by over 16% annually; the white paper also gives a substitution-economics anchor — welding robots have an annual comprehensive cost of about 50,000-80,000 yuan, far below the 120,000-150,000 yuan comprehensive labor cost of a welder, and it forecasts scaled adoption unfolding in three waves, with the timing of the third wave (semi-skilled manufacturing scenarios) depending entirely on the pace of embodied AI technology breakthroughs source
· Coreless motors: Linghu Intelligence disclosed that its coreless motors have a startup mechanical time constant under 28 milliseconds (some products under 10 milliseconds), with energy conversion efficiency above 70% (some exceeding 90%), compared with 20-50% for traditional iron-core motors source (WeChat, CN)
· Inovance Technology: told investors that humanoid robots are a key strategic business, and that it has completed core component development based on its industrial automation expertise, launching frameless torque motors, low-voltage high-power drivers, and linear/rotary actuators source (WeChat, CN)
· South Korea component localization: South Korea is pushing to localize humanoid robot components to close supply chain gaps, with actuator maker Robotis betting on this segment to seize the mass-production window sourcesource
· Dexterous hand transmission: Leadshine Technology's subsidiary Lingqiao Drive & Control published a patent titled "A Finger Module, Manipulator, and Robot," addressing the long-standing challenge of balancing high degrees of freedom with strong grasping force source (WeChat, CN)
Top comments (0)