Today's Highlights
· Tiangong Ultra ran the 400m in 38.15 seconds, beating the human world record of 43.03 seconds
· Shenzhen's robot exports hit RMB 8.58 billion in the first 7 months, up 10.5x year-on-year
· UBTECH's Zigong humanoid assembly line began production a year ahead of schedule, targeting low hundreds of units this year
· AiTu's humanoid robot enters garment factories, reporting a 97% success rate on fabric-piece separation
· California greenlights autonomous trucks over 10,000 lbs for public-road testing
· Octopus Power released a world model, dexterous hand, and data-collection kit in one launch
· JD.com says it will invest RMB 10 billion over three years and open up 10 million hours of real-scene data over two years
Research Progress
World models that only compute physical state predict the wrong next move for people · world-model
The current generation of world models — Sora, Genie 3, JEPA, Marble — encode only objects, positions, occlusion, and motion in their state space, with no representation of "what a person believes happened." The paper's example is brief: if a cup is moved into a cabinet while someone is out of the room, a purely physical model sees nothing wrong with the scene, yet cannot compute where that person will look for it when they return. The authors fold belief, attention, intention, emotion, and social norms into the state and propose a Mental World Modeling framework, splitting every action into a physical carrier (speaking, pointing, grasping) and a mental payload (comforting, deceiving, refusing). The accompanying implementation, MENTIS, requires no additional training: it parses the scene, renders a first-person view, decomposes the action, runs parallel dual-state inference, and scores outcomes by physical plausibility and social appropriateness. For service robots and collaborative agents, this missing layer is a part of the current world-model boom that few have addressed.
Hao Fei, Yiran Zhao · arXiv 2607.27201 source · Analysis: the-decoder source
Open Source · Tools · Benchmarks
· RoboColiseum: a standing simulation benchmark platform built jointly by universities, research institutions, open-source communities, robotics companies, and model developers has launched, with over 40 teams from China and abroad already participating in the closed beta. The platform breaks "task capability" into four dimensions — instruction following, spatial understanding, disturbance adaptation, and general manipulation — spanning 78 high-fidelity simulation tasks, logging failure points at each sub-step, and using domain randomization plus train/test set separation to block models from "memorizing answers." The platform claims sim-to-real correlation of 89.5%, with cross-platform variance for the same model held under 10%, and a fastest registration-to-score time of 30 minutes. Baseline results and training code for ACoT-VLA, π0, π0.5, and GR00T were released alongside. ⚠️ Single-source claim Source: Southern Metropolis Daily source
Funding and Deals
General Intuition (USA) | New funding round in talks | ~$6 billion valuation · world-model ⚠️ Reported, unconfirmed
The company trains models on gameplay footage uploaded by players, and its valuation has jumped from $2.3 billion to $6 billion in two months. It closed a $320 million Series A this past June, led by Khosla Ventures with participation from Jeff Bezos and Eric Schmidt, at a post-money valuation of $2.3 billion at the time; its seed round in October 2025 was $133.7 million, during which it turned down a $500 million acquisition offer from OpenAI. It does not build robots; the goal is to turn millions of hours of gameplay footage into a general-purpose foundation model that any robotics company can call on. Source: Huxiu source
Linsight Technology | New round | Nearly RMB 500 million · hardware
Jointly led by several state-owned investment platforms from Anhui province and Hefei city. The company spun out of the iFlytek (Chinese AI company) system, initially making smart-voice and AIoT chips with cumulative shipments exceeding 150 million units, and is now shifting its focus to edge-side large-model inference for AI PCs, smart cockpits, and robots. Its Nebula series uses a domain-specific-architecture (DSA) approach, stacking 3D DRAM wafers with hybrid bonding to break through the memory wall, with sample chips planned for 2027 targeting 100 tokens/s at 5–15W power draw. Going from "understanding commands" to "understanding the world" requires a different chip design logic; the 150-million-unit shipment figure is a foothold, not a moat. Source: Sohu source
Yin Fangming → Unitree Robotics | Angel round, retrospective | RMB 2 million for a 15% stake · humanoid ⚠️ Estimated figures
In 2016, when Wang Xingxing (Unitree founder) was struggling to raise funds, Yin Fangming approached him proactively and wired the money before any agreement was even signed. At Unitree's IPO price, this investment's total return is estimated at roughly RMB 280 million, over 140x. He did not hold the stake the entire time: he first cashed out in 2018, reducing his stake from 15.0% to 11.4% for proceeds of RMB 1.3106 million; in 2020 he transferred his entire registered capital for RMB 14,300 to Tianjin Junwan Hongyi, in which he himself holds roughly 16.6%. Had he held on to today, the RMB 2 million stake would correspond to a valuation of about RMB 1.848 billion. ROOBO, the company he himself founded, took a different path: in November 2025 it was placed under enforcement by the Beijing No. 1 Intermediate People's Court, with an enforcement target of RMB 83.655658 million. On July 21 this year, Yin Fangming became chairman of Galbot (Chinese embodied-AI startup), with former chairman Guo Xiaoliang stepping down. Source: Ifeng Finance source
Emesent (Australia) | Government CRC-P grant | $2 million · autonomy
The funding comes from the Australian government's Cooperative Research Centres Projects (CRC-P) scheme and will be used to rebuild Emesent's navigation technology into a modular API, forming an open autonomy platform called Cortex AI aimed at GPS-denied environments such as mines, tunnels, and building interiors. Queensland University of Technology is contributing scene-understanding algorithms, with EPE handling field validation. According to Emesent, building GPS-denied navigation from scratch takes more than five years and millions of dollars; the platform aims to remove that barrier and sell the technology to hardware makers and OEMs. Source: Rivista GEOmedia source
Commercialization and Deployment
UBTECH's Zigong humanoid assembly line begins production a year ahead of schedule · humanoid
A nearly 100-meter-long production line began trial production last week and moved to full production this week — originally scheduled for 2027. The multimodal data-collection testing center downstairs went live in January this year, where 150 humanoid robots have been generating data; the company sold its first data order last month, prompting the decision to move the assembly line up by a year. The line currently mainly assembles the Cruzer S2 transport model, with Walker S2 also set to come online; assembly is largely semi-automated, with AGVs handling material delivery while core steps like zero-point calibration, binocular calibration, and stress testing remain manual, allowing flexible switching between models. Bai Mingyuan, head of market operations at the Zigong facility, has set this year's output target in the low hundreds of units, with twenty to thirty skilled workers rotated in from Shenzhen and Liuzhou training local new hires. UBTECH's industrial humanoid production capacity surpassed 1,000 units in 2025, with plans to raise annual capacity to 10,000 units in 2026. Source: Sichuan Online source
AiTu humanoid robots enter garment factories, taking on flexible pre-sewing tasks · embodied ⚠️ Vendor claim
Fabric isn't a rigid part. The same fabric piece bends, folds, and slips once picked up, its shape constantly changing. AiTu (Chinese embodied-AI startup) has put dual-arm humanoids into garment production lines to handle fabric loading/unloading, cut-piece separation, and alignment, along with pre-sewing actions coordinated with template machines, pocket-attaching machines, and interlining-fusing machines. The robots run a domain-specific sewing VLA model and work directly with real components such as cuffs, collars, and pockets. The company's figures: under a single product category and single fabric condition, cut-piece separation success rate is 97%, with downstream sewing pass rate above 98%; the equipment has entered Chinese garment factories for production-line deployment and pilot-scale trial runs. Customers' own calculations are more granular, with an estimated payback period of about 18 months based on equipment efficiency, tied to specific products, fabrics, and test conditions. The error-handling approach is described plainly: the robot retries and self-recovers when possible, skips tasks it temporarily can't complete, and calls in human intervention when judgment is required. Source: Sohu source
Western Australia's largest avocado packing plant installs 9 robots, cuts temp workforce in half · industrial
Avocado Collective installed nine Japanese-made robots at its plant in Manjimup, Western Australia, handing off scanning, palletizing, and pallet transport entirely, at a cost of AU$17 million as part of a broader AU$20 million automation investment. Operations manager Craig Boyce said the machines replaced 40 to 50 workers, "basically everyone who was hand-stacking pallets, everyone scanning, everyone driving forklifts." Three-quarters of the plant's operations are now automated, and weekly capacity has jumped from 1 million kilograms to 2.52 million kilograms. Around 400 local workers come through the federal Pacific Australia Labour Mobility scheme, and unions and labor researchers disagree on how to assess the tradeoff. Source: ABC News source
Qiyuan Q1 opens pre-orders, first shipments planned for September · humanoid ⚠️ Company projection
Shangwei Qiyuan's personal robot Qiyuan Q1 has entered pre-order, with the company saying the first batch will ship in September. The robot was unveiled on December 31 last year by Peng Zhihui (well-known Chinese robotics engineer, "Zhihuijun"), stands about 0.8 meters tall, and folds down to fit into a 30–35-liter backpack, marketed for both portable research-lab use and as a home companion. Personal robots have so far been stuck at the demo-only stage without reaching real sales; how many units actually ship in the first batch is the only number worth watching. Source: Shanghai Observer source
Lucid and Uber bring robotaxis to Houston · autonomy
The two companies' next stop is Houston. Source: InsideEVs source
Industry Developments
Tiangong Ultra runs 400m in 38.15 seconds, nearly 5 seconds faster than the human world record · humanoid
On the afternoon of the 23rd, in the large-class 400m final at the second World Humanoid Robot Games, Tiangong Ultra from the Beijing Humanoid Robot Innovation Center took first place in 38.15 seconds; the men's 400m human world record is 43.03 seconds. In the small-class event, Tiangong Omni ran 45.66 seconds. At the opening ceremony the day before, the standing high jump record was pushed to 2.8843 meters, versus a best of 0.95 meters at the previous games. R&D lead Guo Yijie attributed the improvement to major upgrades in joint torque and rotational speed, higher motor power, and event-specific algorithm optimization — fast acceleration for the 100m, fast cornering for the 400m and 1500m. He did not overstate the results: "At present, robots' athletic performance has reached a certain height, but turning high performance into high stability and high reliability still needs further development." Of the 51 events at this year's games, 30 are competitive events and 21 are scenario-based events; according to event technology partner Huawei, more than 40% require the robots to complete tasks fully autonomously. The scene right after the finish line made a similar point: the organizing committee had laid thick padding a few meters past the finish line, and the robot ran straight into it. Yu Chao, CEO of Lumos Robotics, told Reuters, "Only when it can work in those end-use scenarios does it have real value. Just running and jumping doesn't improve efficiency."Source: The Beijing News source · Reuters source
Shenzhen's robot exports reach RMB 8.58 billion in first 7 months, up 10.5x year-on-year · industrial
According to Shenzhen Customs data, exports of separately classified robots of all kinds from Shenzhen totaled RMB 8.58 billion in the first seven months of this year, up 10.5x year-on-year. "Separately classified" is a new development this year: under the 2026 Tariff Adjustment Plan, intelligent bionic robots have had their own dedicated HS code, 8479.8970, since January 1, with cleaning robots getting separate codes 8508.1110 and 8508.1910; previously, humanoids, robot dogs, and bionic fish/birds were all lumped under the general industrial-robot tariff category. Related statistics show Shenzhen's robot exports account for roughly a quarter of China's total exports of comparable products, sold to more than 100 countries and regions. Mature categories still dominate the volume: from January to June, Shenzhen's cleaning-robot exports totaled RMB 6.025 billion, against RMB 43.43 billion in citywide home-appliance exports over the same period; nationwide, cleaning robots and intelligent bionic robots combined exported RMB 18.09 billion in H1. Humanoids remain at the stage of relying on scenario innovation to find incremental overseas/international demand. Shenzhen's robotics industry had total output value of RMB 242.6 billion in 2025, up 20.56% year-on-year, and citywide industrial robot output grew 74.2% year-on-year in Q1 this year. Source: Caijing source
California allows autonomous trucks over 10,000 lbs to test on public roads · autonomy
Updated rules from the California DMV allow autonomous trucks weighing over 10,000 lbs to test on public roads, with two companies already receiving the state's first such permits. For now, every vehicle must still have a safety driver, and companies must pass strict testing and reporting requirements before they can apply to operate driverless. Supporters make their case around port and agricultural freight, while labor groups have already sued over safety and employment impacts. California is the largest freight corridor in the US and has been one of the states longest holding out against heavy-duty autonomous trucks. Source: USA TODAY source
Octopus Power launches a world model, dexterous hand, and data-collection kit together · world-model ⚠️ Vendor claim
Octopus Power (Chinese robotics startup founded in January this year) simultaneously unveiled the world model SYNWorld 0.5, dexterous hand OctoH-Hand, and data-collection kit OctoSense at WRC. Founder Du Dalong says that in some scenarios, the world model's evaluation of policy performance now matches real-robot evaluation with over 80% consistency; the dexterous hand is currently in small-batch production, with mass production planned by year-end. Du Dalong was Horizon Robotics (Chinese AI chip company)'s employee No. 6 and co-founder/CTO of Jianzhi Robotics; the company closed nearly $50 million in a first round less than two months after founding, with roughly RMB 1 billion raised to date. Source: 21st Century Business Herald source
At a JD.com industry event, humanoid robot transaction volume up 9x year-on-year · adjacent ⚠️ Company projection
JD.com held a robotics industry event during WRC, simultaneously launching an industry-education co-creation program, a robot components industry development alliance, and a robotics strategic upgrade plan. The verifiable existing figures: JD's self-operated platform has partnered with over 200 robotics brands, its European Joybuy platform carries over 50 Chinese brands, and this year JD's humanoid robot transaction volume is up 9x year-on-year, with exoskeleton products up 8x. The pledged figures are a separate matter: RMB 10 billion in resources over the next three years, and opening up over 10 million hours of real-scene data within two years. JD's head of intelligent robotics business disclosed at the event that primary-market investment in embodied AI has exceeded RMB 90 billion, up 5x year-on-year and 1.6x last year's full-year total. Wang Xiaogang, chairman of Daxiao Robotics, pointed to data as the bottleneck at the same event, arguing the industry must shift from scattered data collection to standardized, environment-based collection platforms. Source: Kuaikeji source
Beyond the booths: multiple WRC commentary pieces say most embodied-AI commercialization orders aren't real · embodied ⚠️ Compiled claims
A Huxiu field-observation piece juxtaposes two things: hype around world-model concepts, and the claim that most embodied-AI commercialization orders aren't real; another more widely circulated industry piece puts the figure directly at 80%. These claims come from unnamed industry insiders with no verifiable statistical basis, offering only rough scale and direction. It does line up with another set of publicly disclosed figures from the event: in H1, 60% of humanoid shipments went to performances and research, with fewer than 20% entering core production settings (previously reported). Source: Huxiu source
Hardware · Supply Chain
· Dexterous hand pricing: per CCTV Finance, the average price was RMB 50,000 per unit in 2024; this year it can generally be brought under RMB 10,000, with six-DOF products as low as RMB 6,000–8,000. Yicai's on-site reporting gives a wider range, from RMB 500,000–1,000,000 down to tens of thousands of RMB. The Kuaikeji cailianshe (Chinese financial-tech outlet) STAR Market Daily tracked an even wider price band at WRC — Yuequan Bionics simultaneously released a sub-RMB-10,000 dexterous hand, Xinuo Future's rope-drive-plus-direct-drive hybrid five-finger hand sells for RMB 128,000, while some high-end products still run from hundreds of thousands to over a million RMB. Eight years ago, at the first China International Import Expo, a pair of German dexterous hands cost roughly RMB 2 million. sourcesourcesource
· Dexterous hand sales and funding: GGII (Chinese industrial research firm)'s "2026 H1 China Dexterous Hand Industry Survey Report" puts 2025 Chinese dexterous-hand sales at roughly 19,200 units, projected to reach 70,200 units in 2026 and potentially surpass 430,000 by 2030; STAR Market Daily, citing public data, reports over RMB 25 billion in total funding for Chinese dexterous-hand startups in H1 2026. sourcesource
· Tactile sensors: per industry figures, tactile-sensor penetration in commercialized humanoid dexterous hands has risen from 20% in early 2025 to over 60%, with roughly RMB 3 billion in cumulative sensor-sector funding in H1, up 160.87% year-on-year. Fingertip 3D force solutions can now sense both normal and shear force simultaneously, with array-based tactile sensing reaching a detection floor of 0.005 N. Paxini (Chinese tactile-sensing company) brought the world's first multi-dimensional plantar tactile sensor, PX-FOOTRIX, and its 4th-generation 6D tactile sensing chip, GEN4 FUSE, to this year's event. sourcesourcesource
· Infineon: co-developed a dexterous-hand tactile solution with Sensen Technology (Chinese sensor company), combining CAPSENSE capacitive sensing with flexible capacitive sensors — a single robotic hand carries 6 PSOC 4 chips covering fingers and palm by zone; Infineon also showed a humanoid head platform using three CMOS radar chips for 360° monitoring. In fiscal Q3 2026, Infineon has signed or is negotiating multi-year capacity-reservation agreements with several major AI customers, together representing €7–9 billion in potential revenue. source
· Orbbec: released a platform-agnostic family of data-collection hardware — first-person EGO, handheld UMI, wrist-mounted near-field WristCam, and a multi-device sync Hub. The company claims continuous recording of over 37 hours under certain conditions with zero RGB/IMU frame drops, and multi-camera hardware sync error under 1 millisecond. Founder Huang Yuanhao lays out the business logic plainly: customers can swap suppliers anytime if you only sell hardware, but data is different — "data is inherently tied to the model, so its importance and stickiness are completely different." ⚠️ Vendor claim source
· Yuequan Bionics: its Yingshou Y-Hand M0 weighs 440 grams with 16 degrees of freedom and a max fingertip output force of 100 N, able to grip both a wrench/drill and vegetables/flowers; its self-developed slotless brushless coreless-cup servo motor was released in March this year and entered mass production in July. Micro-motors are the shared bottleneck for both cost and performance in dexterous hands. source
Top comments (0)