Today's Highlights
· Google's Gemini Robotics 2 for the first time controls everything from bipedal locomotion to a 22-DOF five-finger hand in one model
· China's National Development and Reform Commission: for every 10 humanoid and quadruped robots sold globally, 8 are made in China
· LimX Dynamics closes a nearly $200 million Pre-IPO round at a post-money valuation of RMB 15 billion
· Unitree's IPO prospectus responds to the FCC: models currently on sale are already certified, but new models may be unable to sell in the US
· Dexmal (Chinese embodied-data startup) open-sources ACE-Data-0, 17 million frames of real-home multimodal data
I. Research Progress
ACE-Data-0: Turning two real houses into synchronized recording studios, capturing all six modalities in one pass · benchmark
A longstanding problem in embodied data is fragmentation — first-person view, full-body motion, hand manipulation, object state, sound, and touch are scattered across different datasets, so the perception-action loop never sees more than half the picture. The ACE capture engine turns real homes into spatially calibrated, time-synchronized recording venues, resolving tabletop-scale hand-object interaction and room-scale full-body movement, recording all six modalities together as a single multisensory stream. The resulting dataset spans 200 task categories, 150 hours, and 17 million frames, with 75,000 interaction episodes from 50 participants; goal-level rather than step-by-step instructions preserve natural behavioral variation. An accompanying tiered benchmark shows existing SOTA methods still have clear gaps in contact, occlusion, self-motion, and long-horizon reasoning. It ranked highest in community signal among today's papers, with an HF trending score of 33↑.
Yukang Cao et al. · arXiv 2607.28625 source · Open-sourced by: Dexmal (Chinese embodied-data startup)
World Action Planner: Have a VLM propose plans and a world model referee them, outperforming VLA · world-model
The dominant answer over the past year has been to cram planning and control together end-to-end into a VLA or WAM. This paper goes the other way: use a VLM's reasoning ability to first propose an initial action plan, then "imagine" it forward inside a multi-task, pose-and-image-conditioned world model, iteratively refining through optimization and search. The authors report significant improvements over the best current end-to-end VLA and WAM approaches across compositional tasks, novel layouts, and zero-shot generalization — an empirical vote for the world model as planner rather than generator, in a debate that has recurred all year.
Xiangcheng Zhang, Yilun Du · arXiv 2607.27599 source
TacWAM: Letting a world model predict tactile futures, but the action branch isn't allowed to peek · manipulation
World action models generally only predict visual futures, which provide almost no supervisory signal for force, deformation, shear, or slip. TacWAM maps tactile appearance, dense force fields, and deformation flow into a shared latent space, then uses anchored tri-modal attention to separate "current observation anchor / future prediction tokens / action tokens" — future tactile information only participates in training supervision and is invisible to the action branch, preventing information unavailable at deployment time from becoming a training shortcut. Across four contact-rich real-robot tasks it reaches an average success rate of 75.0%, 37.5 percentage points above the strongest baseline.
Lei Jin et al. · arXiv 2607.28391 source
The key to cross-embodiment transfer isn't data volume, it's the intermediate representation · vla
Large-scale imitation learning relies on pooling data across multiple robot embodiments, but true cross-embodiment transfer has remained hard. This work systematically compares "behavior-aligned representations" — object boxes, verbalized actions, end-effector trajectories — in VLAs, finding that end-effector trajectories are especially effective for transfer, with the benefit growing as the prior dataset scales, and that they can also absorb action-free data. In sim-to-real cross-embodiment transfer, real-robot policies pretrained on simulated data show a 28% improvement in task completion.
Ajay Sridhar, Jensen Gao, Jonathan Yang, Jean Mercat, Suneel Belkhale · arXiv 2607.27549 source
FasTac: Packing multispectral photometric stereo and force estimation into an FPGA to build a curved fingertip · perception
Curved tactile fingertips needed for dexterous manipulation must resolve fine contact geometry, distinguish normal from shear loads, and capture transient signals simultaneously — existing curved vision-based tactile sensors struggle to deliver all three in a compact form factor. FasTac integrates multispectral photometric stereo, dynamic convolutional force estimation, and FPGA hardware acceleration together, providing high-speed 3D shape reconstruction and tri-axial force estimation simultaneously — exactly the kind of "manufacturable form factor" device work that's been missing upstream of dexterous hands.
Xiaofan Lu et al. · arXiv 2607.28416 source
UniCross: Grasping, transfer, in-hand rotation, and in-hand translation are actually the same problem · manipulation
Viewed through hand-object relative motion, dexterous manipulation requiring stable grip can be decomposed into four basic skills — grasping, relocation, in-hand rotation, and in-hand translation — which the human hand freely combines. This work unifies all four skills under the same state space, action space, and objective function, suggesting that different dexterous skills are simply different instances of the same task formulation — a structural simplification direction relative to today's practice of training each single skill separately.
Hui Zhang, Julian Ferchow, Jie Song, Mirko Meboldt · arXiv 2607.28198 source
ODEWorld: The physical world is continuous, but world models keep working frame by frame · world-model
Existing world-modeling paradigms are largely constrained to discrete-time prediction, which limits how efficiently they capture physical dynamics. This paper proposes Physical-Time Flow, learning a continuous latent velocity field that operates in physical time; the authors say it simultaneously alleviates the long-standing representation collapse problem in latent-space world models.
Dongxiu Liu, Haoyi Niu, Peng Cheng, Yuan Gao, Xirui Kang · arXiv 2607.27924 source
PAC-MAN: Humanoid robots dodge incoming balls, with control barrier functions managing whole-body safety · locomotion
Combines the safety guarantees of control barrier functions with reinforcement learning, paired with deployment-grade onboard perception rather than god's-eye-view state, for whole-body humanoid ball-dodging. This kind of "whole-body safety under limited perception" setting is closer to real deployment conditions than enforcing safety constraints under perfect state feedback.
Lizhi Yang, Junheng Li, Aaron D. Ames (Caltech) · arXiv 2607.28623 source
Other papers today: EgoGenesis synthesizes controllable first-person manipulation videos using online anchored-projection memory to augment scarce real-robot data (arXiv 2607.28243 source); SIDO uses counterfactual action augmentation so policies trained only on static-object demonstrations adapt to moving objects at test time (arXiv 2607.27890 source); RedFlow rewrites failure experience into action-level corrective supervision fed to flow-matching VLA policies (arXiv 2607.27782 source); QuantWAMs redesigns post-training quantization calibration for the closed-loop execution characteristics of world action models (arXiv 2607.28405 source); Articulated Object Reconstruction reconstructs articulated objects' geometry and kinematic structure from static closed-state observations alone (arXiv 2607.27749 source); FA-RDP keeps diffusion policies multimodal before contact and switches to high-frequency reactive behavior after contact (arXiv 2607.28596 source); X-NavDP uses group Q-score reweighted matching for cross-embodiment post-training of navigation diffusion policies (arXiv 2607.28560 source); CS-JEPA has each robot in a swarm predict the same shared future using only local observations (arXiv 2607.28443 source); DexDirect lowers the barrier to dexterous demonstration collection using direct kinesthetic guidance plus visual hand retargeting (arXiv 2607.27784 source); RaDiVe improves 4D mmWave radar odometry robustness with a velocity-discrepancy-point uncertainty model (arXiv 2607.28045 source); SemAnCorr establishes dense semantic correspondence training-free for zero-shot manipulation-skill transfer (arXiv 2607.28382 source); RoboBRIDGE wraps a monitor-perceive-plan-control orchestration layer around off-the-shelf VLAs (arXiv 2607.27881 source); a survey systematically reviews how foundation-model priors are injected into hand-object interaction reconstruction, generation, and embodied transfer (arXiv 2607.28394 source).
Open Source · Tools · Benchmarks
· LingBot-World 2: Ant Group's LingBot (Ant Group's embodied AI unit) open-sources a next-generation real-time interactive world model, pushing continuous lossless generation from roughly ten minutes in version 1.0 to the hour scale, holding a stable 60fps at 720p resolution; the action library expands to complex interactions like attacking, casting, and shooting, and supports typed text injected as events that rewrite the world. A new Agentic Harness lets the world keep evolving even without human operation. The technical report bakes causality (state depends only on history and current input) into the architecture, uses MoBA hybrid attention masking to suppress long-sequence drift, then applies a two-stage teacher-student distillation to get both "long" and "fast." source (WeChat, CN)
· WALL-WM: A world action model open-sourced by AgiBot (Chinese humanoid startup), replacing fixed-duration action chunks as the training unit with "action events" (approach, contact, grasp, lift, move, place); the video tower reuses weights from the Wan family of text-to-video models, with an action DiT cross-attending layer by layer to avoid overwriting visual semantic priors. source
· ASIMOV-Agentic: A safety benchmark released by Google DeepMind alongside Gemini Robotics 2, evaluating a robot's ability to refuse unsafe actions, determine a task cannot be completed safely, and proactively request human intervention. source
· WorldScape 2.0: Manifold AI releases full-stack technology, arguing that world-model scaling should shift from "rebuilding a model for each new control" to "plugging a new control into the same world." ⚠️ Vendor claim source (WeChat, CN)
· World-model video generation capability assessment: Officially launched by China Academy of Information and Communications Technology, setting evaluation criteria for multimodal perception, spatial modeling, and causal reasoning capability. source (WeChat, CN)
II. Funding & Deals
LimX Dynamics ｜ Pre-IPO round ｜ Nearly $200 million ｜ Post-money valuation RMB 15 billion · humanoid
The round was led by IDG Capital, Lens Technology (Chinese electronics component maker), GGG Group, and Redstone VC, Huashan Capital, and Hefei Binhu Industrial Development Group, among other strategic investors, with UAE-based Lunate (Emirati investment firm) adding across multiple consecutive rounds and existing shareholders Oasis Capital, Cornerstone Capital, Nanshan Strategic Emerging Industries Fund, Shangqi Capital, and NIO Capital all over-subscribing. This makes it the second time in six months this member of Shenzhen's "eight humanoid heavyweights" has raised — nearly $400 million total. The structure is more notable than the amount: nearly 70% of the round came from European, Middle Eastern, and North American institutions, unusual among Chinese embodied-AI companies and pointing toward overseas channel access rather than pure capital; Lens Technology's involvement fills in precision assembly and mass-production quality control — the company says its full product line has already secured orders for several thousand units in total, over half from overseas, with the first several hundred units already shipped, the kind of delivery that requires consumer-electronics-grade manufacturing systems. Founder Zhang Wei is a professor at Southern University of Science and Technology who has researched humanoid robots for over a decade; the company completed its shareholding restructuring in March this year and is steadily advancing toward an IPO. Source: PEdaily source (WeChat, CN)
LiberAI ｜ Pre-A+ round ｜ Several hundred million RMB · world-model
Jointly invested by 360's strategic investment arm, China Development Bank Capital, Sinovation Ventures, CMC Capital, and Binfu Capital, with existing shareholder Sequoia China adding across four consecutive rounds since the angel round; over the past three months alone it has secured consecutive backing from Sequoia, ZhenFund, Meituan Long-Zhu Fund, and Shunwei Capital — this is the fifth round completed by the company, founded only in late 2025, within 8 months. Its technical approach bets on "UMI human data + world models": an internet human-video base layer, a proprietary data glove as intermediate layer, and real-robot data at the top of a "data pyramid," with human video volume 1-2 orders of magnitude larger than glove data, and glove data in turn 1-2 orders of magnitude larger than real-robot data. The company says its 20-DOF dexterous hand has completed the full training-to-deployment pipeline, capable of two-handed banana peeling and bottle-cap twisting, with its proprietary data glove now in mass production. Founder Liu Songming is a post-2000 Tsinghua Special Prize winner who studied under Zhu Jun and previously led the release of the RDT-1B foundation model. Source: Tech Capital Circle source (WeChat, CN)
Zhongke Yuandongli ｜ Series B2 ｜ Several hundred million RMB · industrial
Invested exclusively by Chery Holding's Anhui Chery Smart Technology unit, marking the company's largest disclosed funding round to date, with proceeds earmarked for R&D on smart electric agricultural machinery and embodied-AI agricultural robots, scaled delivery, and overseas expansion. Founded in 2019 as a spin-off from the Chinese Academy of Sciences' Institute of Microelectronics, its "Wantu" line of smart electric agricultural machinery has already been commercialized with centimeter-level operating positioning accuracy; its autonomous-operation system can be retrofitted onto ordinary farm machinery and has completed over 1 million mu (about 165,000 acres) of cumulative standardized autonomous operations across more than 20 provinces and regions. Its chief scientist is Li Deyi, an academician of the Chinese Academy of Engineering. Source: Tech Capital Circle source (WeChat, CN)
Quantum Dynamics ｜ Seed round ｜ Over RMB 100 million · embodied
Jointly invested by Yunqi Capital and SenseTime's Guoxiang Capital. Founder and CEO Li Qiang spent 17 years at Alibaba, having served as CTO of Cainiao Group, CTO of Alibaba's international digital commerce division, and general manager of Cainiao's unmanned vehicle unit. The company is entering through logistics: starting in 2026 with the three most labor-intensive roles — restocking/shelving, picking, and packing — for B2C and small-B2B warehouses, beginning with SKU-rich, well-organized scenarios like cosmetics and pharmaceuticals in Chinese e-commerce warehouses, before expanding to more complex courier-network sites after 2027. Its technical approach advocates an embodied-native World-Action Model that jointly understands vision, action, proprioceptive state, and touch, with a lightweight World Model Corrector that catches and retries actions before they fail. Source: Zhidongxi source
Zhenlan Instrument acquires 66% of Wuhu Mokka ｜ RMB 198 million · industrial
Zhenlan Instrument (301303) announced it will acquire a 66% stake in industrial-robot developer and manufacturer Wuhu Mokka for RMB 198 million in cash from its own funds. It has already secured 56.83%, with the remaining 9.17% subject to a public listing-and-transfer process through a property-rights exchange, carrying some uncertainty. The target company reported revenue of RMB 52.88 million and net profit of RMB 3.71 million for January-May 2026, with total shareholders' equity appraised at RMB 300 million as of May 31. Source: Sina Finance source
III. Commercialization & Deployment
Komatsu partners with AIM: autonomous bulldozers and excavators already at work on US customer sites · industrial
Komatsu and its subsidiary EARTHBRAIN have entered a strategic partnership with AIM Intelligent Machines to provide autonomous construction solutions for the Japanese and US markets. The two sides had previously advanced joint technical validation, and the partnership has now entered commercial deployment in the US market — Komatsu's autonomous machinery is already running at customer sites, with Japan deployment planned for 2027. The solution feeds construction plans and target terrain data generated by Smart Construction into AIM's physical AI platform, letting bulldozers and excavators determine their own construction methods and travel routes and execute independently. One notable point: AIM's technology can be retrofitted onto existing equipment, meaning customers don't need to wait for new machines and can adopt it in batches — a practical way to lower the barrier for capital-intensive earthmoving fleets. Source: International Mining source
Hai Robotics lands local racking supplier and integrator in Australia, global project count passes 1,700 · industrial
Hai Robotics has established a strategic partnership with Australian racking supplier Dexion Victoria and systems integrator Dawson Robotics, targeting high-density warehousing demand in the local market. Dexion Victoria will complete installation of a HaiPick Climb demonstration system at its new warehouse in Dandenong South, Victoria in August 2026, serving as a live validation platform for local customers. The company established its Australian subsidiary in 2022 and has now deployed over 1,700 automation projects globally, with operations spanning more than 40 countries and regions. Source: MHD Supply Chain source
Malaysia's first autonomous-logistics remote operations center launches, Cyberjaya opens a 21km L4 test route · autonomy
A joint effort by the Selangor Information Technology and Digital Economy Corporation, Autonomous Logistic Solutions, and Cyberview has launched Malaysia's first autonomous-logistics remote operations center at CoPlace 9 in Cyberjaya, enabling the local MyAV test route to support L4-level autonomous logistics vehicles. The route has been approved by the Ministry of Transport as a controlled public-road testing site, and over the coming months will run tests in real mixed-traffic conditions along a route spanning 21 kilometers with 21 stops. The operations center integrates real-time video streaming, telemetry data, route status, and system diagnostics, with dedicated staff monitoring throughout and coordinating ground support in emergencies; data collected during testing will feed into Malaysia's national standard-setting process for autonomous logistics. Source: Oriental Daily Malaysia source
Geely Farizon's Xinghan H completes a 1+4 commercial-vehicle platooning trial, lead truck manned, following trucks developed to L4 · autonomy
The Xinghan H 6x4 pure-electric tractor completed a multi-truck platooning autonomous-driving trial at the Shandong Expressway Intelligent Connected Vehicle Testing Base's Laiwu section, achieving full 1+1 to 1+4 platoon capability. Testing covered one-key platoon formation, coordinated lane changes via turn signal, minor lane-internal obstacle avoidance, other-vehicle cut-in/cut-out, and on/off-ramp merging; the platoon uses an L2++ navigation-assist lead truck with following trucks developed to L4. The commercial logic is straightforward: tight platooning reduces wind resistance and energy consumption, following trucks don't need a safety driver — freeing up labor costs — and removing the cab frees up cargo space and boosts effective payload simultaneously. ⚠️ Closed test-base conditions, with expansion to real open roads, western desert trunk routes, and port collection/distribution scenarios to follow. Source: NetEase source
Deep Robotics' wheeled-legged robot enters Serling Tso to film Tibetan antelope, establishing routine follow-shooting · adjacent
During the Tibetan antelope migration season, a Deep Robotics (Chinese quadruped/wheeled-legged robot maker) Lynx M20 wheeled-legged robot fully documented migration, calf-rearing, and foraging footage along the shore of Serling Tso Lake. The unit's IP66 rating withstands sudden plateau rain, snow, and hail, with a -20°C to 55°C wide temperature range suited to the region's persistent low temperatures; it uses a 96-line LiDAR and depth camera to autonomously perceive lakeside meadows, permafrost undulation, and rocky slopes. Its bio-inspired quiet operation lets it observe at extremely close range without disturbing the herd. Last year the company's X30 unit at Hoh Xil took a "robot-antelope" disguised-integration approach; this year the Lynx M20 shifts to long-duration routine follow-shooting, validating practicality in extreme high-altitude conditions rather than a one-off demonstration. Source: Huanqiu source
MicroPort's Toumai MT-1000 Pro surgical robot installed at a hospital in Madurai, India · embodied
Following a sharp rise in Chinese surgical-robot exports in the first half of the year, Meenakshi Super Speciality Hospital in Madurai, India has introduced the next-generation Toumai MT-1000 Pro surgical robot from MicroPort (Chinese medical-device maker), covering gastrointestinal, hepatobiliary, urological, gynecological, thoracic, breast, thyroid, and pediatric surgical procedures. The system uses ultra-low-latency master-slave control, wristed instruments, and laser-assisted smart docking, with surgeons operating from a dedicated console while the system filters out hand tremor. Hospital chairman Dr. S. Gurushankar said the platform also opens up new possibilities for remote surgical collaboration. Source: Digital Health News source
IV. Industry Developments
Google DeepMind releases Gemini Robotics 2, for the first time letting one model go from feet to fingertips · world-model
Robotic AI has largely been an "upper-body system" until now: arms controlled by one controller, legs by another, stitched together through handoffs that break down whenever the environment or payload changes. Gemini Robotics 2 pushes this to the whole body — legs, torso, and both arms down to a 22-DOF five-finger hand are coordinated by a single end-to-end policy, generating whole-body control signals directly from camera images and natural-language instructions. The demo runs on Apptronik's Apollo 2: told to "put the watering can in the green bin on the bottom shelf," it walks over on its own, picks it up, steps forward, and crouches to place it down; fitted with a SharpaWave five-finger hand it can also tie off a trash bag and screw in a lightbulb, while still keeping a parallel gripper option for industrial insertion and packing tasks. Three models were released together: the flagship VLA handles action, Gemini Robotics ER 2 handles embodied reasoning and task management spanning several minutes and for the first time supports different robot models communicating and collaborating with each other (the demo shows an Apollo 2 and a Franka F3 Duo splitting the same task), and On-Device 2 runs entirely on the robot's local hardware for unstable industrial-site networks. Google says adapting to new hardware takes only a few hours and fewer than 200 training samples to transfer to a new dual-arm platform. ⚠️ Vendor claim. One point in the strategic positioning worth noting on its own: the partner roster includes Apptronik, Boston Dynamics, and Franka — Google isn't building bodies, it's positioning itself as the "brain supplier for every robot," the opposite approach from Tesla and Figure building their own production lines for hardware. Source: Robotics & Automation News source
NDRC: for every 10 humanoid and quadruped robots sold globally, 8 are made in China · humanoid
At a press briefing, China's National Development and Reform Commission disclosed that for every 10 humanoid and quadruped intelligent robots currently sold globally, 8 are made in China. This is the first time official Chinese sources have put this figure this specifically, and it provides a reference point for several recent third-party data points: Morgan Stanley recently raised its 2026 China humanoid-robot shipment forecast to 50,000 units, nearly double its previous forecast of 28,000, and projects China's humanoid market to reach $2 billion in 2026, growing to $15 billion by 2030 with annual shipments of 446,000 units; Omdia's data shows 13,000 humanoid robots shipped globally in 2025, with the top five all Chinese companies, and Figure AI and Tesla ranking seventh and ninth respectively. Embodied AI has been included in China's "15th Five-Year Plan" future industries blueprint. Source: IT Home source
Unitree's IPO prospectus directly addresses the FCC ban: models on sale are already certified, new models risk being unable to obtain certification · humanoid
Following the August 10 subscription date set earlier, the Shanghai Stock Exchange updated four announcements today: preliminary inquiry is set for August 5, with payment deadline August 12; two employee asset-management plans participating in the strategic placement subscribed for a combined RMB 271.5 million, with founder Wang Xingxing personally subscribing RMB 15 million as the largest individual contributor, and several core executives subscribing between RMB 1 million and 9 million each. More notable is the special risk disclosure: in response to the FCC's July 28 addition of overseas-manufactured advanced robots to a regulated list, Unitree disclosed that its main models currently on sale — the G1, H2, R1 humanoids and Go2, B2, A2 quadrupeds — have all already obtained FCC certification, so the policy change does not affect continued sales of existing products in the US; however, future new models developed afterward will risk being unable to sell in the US unless granted an exemption or conditional approval. This risk carries real weight in the revenue mix: overseas revenue exceeded 40% of the total in each reporting-period year, with the US market share at 18.39%, 19.54%, and 13.30% respectively. On performance, the company expects revenue of RMB 1.052 to 1.128 billion for January-June 2026, up 35.62% to 45.41% year-on-year, but net profit excluding non-recurring items declined year-on-year, mainly due to a substantial increase in R&D spending and large first-quarter sales expenses. Source: Toutiao Tech source (WeChat, CN)
FCC's robot list also covers robot vacuums, reach far broader than humanoids · adjacent
FCC media relations director Katie Gorscak confirmed to The Verge that robot vacuums are also within the scope of the regulation. Per the FCC's definition, restricted devices are software-controlled robots that are ground-traveling, weigh over 4.4 pounds including their base, have environmental-sensing capability, and carry wireless connectivity — which also covers robotic lawnmowers, sidewalk delivery robots, and warehouse robots that move packages. Already-approved models on sale are unaffected; new products will need to apply. Reportedly, companies seeking an exemption must demonstrate local manufacturing or commit to investment in US manufacturing. The impact is direct: IDC analyst Jitesh Ubrani noted that the top five robot-vacuum brands — Roborock, Ecovacs, Dreame, Xiaomi, and Yunjing — are all Chinese companies, Dyson also relies on Chinese manufacturing, and following iRobot's bankruptcy the Roomba brand is now owned by Chinese company Picea. Even Matic, which assembles in California, says it may still need to apply for an exemption because its share of US-sourced components does not meet the 65% threshold. Source: Dexerto source
IDC: China's industrial embodied-AI robot market reached RMB 5.74 billion in 2025, with humanoids accounting for only RMB 2.11 billion · industrial
IDC data shows China's industrial embodied-AI robot market was approximately RMB 5.74 billion in 2025, of which embodied applications carried on industrial-robot platforms made up the majority at approximately RMB 3.62 billion, spanning collaborative robots, composite mobile-manipulation robots, and multi-joint robots, already in use for loading/unloading, quality inspection, grinding and repair, flexible assembly, and material handling; new platforms represented by humanoids accounted for approximately RMB 2.11 billion, with applications still concentrated in demonstration production-line deployment, scenario validation, and POC testing. This breakdown offers a useful reference point for the current valuation debate — the segment actually making money remains the "hardware + software + services" full package around mature industrial-robot categories. IDC's user survey also found that over 80% of manufacturing enterprises hope to recoup project investment within two years. Source: Zhitong Finance source
AgiBot pays a company-wide mid-year special bonus, departed employees get a pro-rated payout too · humanoid
AgiBot (Chinese humanoid-robot startup) announced a "6+1" half-year double-pay policy at its mid-2026 all-hands meeting, giving all formal employees an extra month's fixed salary; employees who left between January 1 and June 30 will also receive a payout pro-rated by days worked, with the company saying its first-half performance targets were exceeded. Using contribution during the work period rather than employment status at payout time as the criterion sparked considerable discussion given the industry's broadly tightening benefits environment. In the same period, the company announced ten partners in the UK and Ireland markets in London, advancing its European expansion; its Hong Kong listing process is still underway. Source: Red Star Capital Bureau source (WeChat, CN)
ByteDance releases Seedance 2.5, XPeng, AgiBot, and XCMG among first adopters · world-model
ByteDance released its video-generation model Seedance 2.5, extending single-generation length to 30 seconds; Volcano Engine announced that XPeng Motors, AgiBot, and XCMG (Chinese construction-machinery maker), among others, will be among the first to integrate it. Video-generation models moving toward embodied-AI and heavy-machinery customers points to one thing: synthetic data and scene generation are becoming a routine procurement item in these companies' training pipelines, not just a lab toy. Source: Ifeng Tech source
A humanoid robot backed by Qualcomm falls flat on stage at Computex · humanoid
At the Computex 2026 keynote, a humanoid robot powered by Qualcomm chips suddenly collapsed on live stream, and the video quickly spread. Beyond the spectacle, it's a useful reminder: what stands between stage demos and deliverable productivity is exactly stability and failure recovery — the same reason Gemini Robotics 2's release today specifically called out failure recognition and retry as a capability, alongside a companion safety benchmark. Source: Interesting Engineering source
Hardware · Supply Chain
· Weihong Electronic: First-half dexterous-hand joint-motor/module shipments surpassed 50,000 units, with products now integrated into roughly 40% of dexterous-hand solutions globally and about 80% in China; production-line expansion targets monthly output of 80,000 units by end of 2026. First-half revenue was RMB 286 million, up 10.06% year-on-year, with net profit attributable to shareholders of RMB 6.0243 million, down 79.4% year-on-year (mainly due to non-recurring factors). ⚠️ Capacity target per company disclosure source
· Future People · Qiyuan: Xiamen Future People Intelligent released an open-source dexterous hand with a 1:1 human-like hand proportion, driven by 21 proprietary servo motors, 42 independent degrees of freedom, and 47 total ranges of motion across the hand, priced under RMB 40,000 — compared to high-end dexterous hands overseas typically priced in the hundreds of thousands of RMB. All finger cavities are left empty with no pre-installed sensors, allowing array tactile sensors, e-skin, or miniature LiDAR to be integrated through standard interfaces for validation. ⚠️ Single-party claim source
· Dexterous-hand shipment forecast: Industry forecasts put 2025 dexterous-hand shipments at roughly 30,000-40,000 units, rising to 100,000 units in 2026 with nearly all new products concentrated in high-DOF designs; 2027 high-DOF dexterous-hand shipments are projected at 100,000-400,000 units, with a neutral estimate of 200,000. ⚠️ Research-institution forecast source
· Midea Group (00300.HK): the company is building new reducer-gear production capacity of 300,000 units. source
· China National Technology: is laying out MCU products targeting robot "cerebellum" control modules, joints, dexterous hands, and encoders. source
Top comments (0)