Today's Highlights
· South Korean conglomerates Samsung, Hyundai, SK and others announced a combined roughly ₩312 trillion (about $195 billion) in southern-region industrial investments on the same day, betting heavily on physical AI manufacturing — Samsung alone will invest ₩60 trillion to build a physical AI manufacturing cluster in the Yeongnam region, plus another ₩19 trillion for a robotics factory in Gumi.
· Industrial robot leader Estun Automation (Chinese industrial robotics maker) plans to fully acquire its collaborative-robot subsidiary Estun Cobot to expand into embodied AI, with the stock hitting the daily limit-up 3 times in 4 days.
· XBOT, a food-service embodied AI company founded by former Xiaomi executive Tang Mu, closed two funding rounds totaling hundreds of millions of yuan (Series A: ¥200 million; Series B: ¥300–500 million).
· Unitree Robotics' (Chinese humanoid robot maker) IPO registration approval ignited China's A-share robotics sector, with over 40 stocks hitting limit-up in a single day (the IPO was approved yesterday; today's move reflects secondary-market trading).
· Autonomous driving company Momenta (Chinese self-driving startup) saw its Hong Kong IPO public offering oversubscribed by more than 414 times, as the Hong Kong listing window for robotics/autonomous-driving companies stays hot.
1. Research Progress
Learn to Move Before You Learn to Act: Task-Agnostic Pretraining for VLA (TAP) · vla
VLA models have long been bottlenecked by scarce expert demonstration data. This paper separates "learning how to move" (motor competence) from "learning what to do" (semantic alignment): it first uses cheap, unlabeled data — including discarded off-task trajectories and autonomous robot play — to learn transferable motion priors via self-supervised inverse dynamics, then aligns those priors to language instructions using very few labeled demonstrations, significantly reducing reliance on expensive demos. It gained HF↑3 on the day, and the approach has broad value for data-constrained settings.
Junhao Shi et al. · arXiv 2607.02466 source
One Demo Is Enough: Real-World Robot Reinforcement Learning (AutoSERL) · manipulation
Real-world RL faces two major pain points: expensive data and the need for continuous human intervention during training. AutoSERL fully automates the intervention process using just a single demonstration, achieving 100% success rate on insertion tasks and greater robustness to positional perturbations — cutting the startup cost of real-world RL down to one demo.
Yuwan Liu et al. · arXiv 2607.01651 source
PhysMani: A Physics-Grounded 3D World Model for Dynamic Manipulation · world-model
Grasping fast-moving targets remains a hard problem for embodied AI, and existing VLA and world models both struggle to jointly achieve precise 3D geometry and physically plausible predictions. PhysMani couples a physics-grounded 3D Gaussian world model with a "future-aware" action policy, so predictions respect both geometry and physics, for dynamic object manipulation in unstructured 3D environments.
Peng Yun et al. · arXiv 2607.01938 source
Bridge-WA: Predicting Only Where and How the World Changes, Not Rendering Full Future Frames · world-model
General-purpose VLA benefits from large vision-language priors, but effective manipulation also requires anticipating action-relevant scene changes; large generative world models that do dense future rollouts are expensive, spending much of their compute on visual details only weakly related to control. Bridge-WA distills a frozen "future-change teacher" into three compact priors (future tokens, change maps, motion-flow maps), lightly boosting task success and progress — a pragmatic route of "predicting less pixels, more causality."
Yongjie Bai et al. · arXiv 2607.02195 source
VT-WAM: A Visual-Tactile World-Action Model for Contact-Rich Manipulation · world-model
Contact-rich manipulation requires instant response to local deformation, pressure, slip, and friction — cues that are sparse or even invisible in vision. Existing visual-tactile policies mostly feed tactile signals directly into action prediction, rarely modeling the dynamics of tactile deformation during action generation. VT-WAM brings tactile sensing into the world model's predictive loop, filling this gap.
Shuai Tian et al. · arXiv 2607.02503 source
Imagining Touch: "Tactile-Informed" Manipulation Without Tactile Hardware · perception
Tactile sensing can substantially improve contact-rich manipulation, but sensors are fragile, need calibration, and are costly to maintain, limiting real-world deployment. This paper proposes imagined tactile representations — at deployment, no tactile sensor is installed, yet the robot still gains the benefits of tactile knowledge, answering the fundamental question of whether tactile benefits are attainable without tactile hardware.
Zhiyuan Zhang et al. · arXiv 2607.01684 source
Guided Action Flow: Adding Inference-Time Q-Guidance to Flow-Matching VLA · vla
Flow-matching VLA generates action chunks through iterative transport, naturally leaving room for "no retraining needed" test-time guidance. This framework keeps a pretrained SmolVLA policy frozen and uses a learned action-chunk critic to guide its reverse flow sampler, improving performance without touching the backbone weights — another example of the low-cost "frozen backbone + inference-time polishing" paradigm.
Liuhaichen Yang et al. · arXiv 2607.02092 source
HEFT: Heavy-Payload Teleoperation for Full-Scale Humanoids · locomotion
General motion tracking/teleoperation is one path to scaling humanoid skills, but most frameworks are validated on small platforms or without real payloads, leaving full-scale humanoids under real load almost entirely unstudied. HEFT learns from noisy VR references using privileged motion guidance, then applies a windowed payload curriculum to progressively add load, achieving robust heavy-payload tracking.
Chenxin Liu et al. · arXiv 2607.02332 source
Other papers today: VLA-Corrector adds lightweight "detect-and-correct" reasoning with on-demand adaptive action horizons to chunked VLA (arXiv 2607.01804 source); The Moving Eye uses a dual-arm setup where one arm operates while the other serves as a moving camera, improving VLA spatial generalization (arXiv 2607.02322 source); WorldSample closes the "real-synthetic" data augmentation loop in real-world RL using a world model (arXiv 2607.02431 source); ACID uses inverse-dynamics consistency to constrain the achievability of intermediate states in world-model planning (arXiv 2607.02403 source); Actuator Reality Shaping shapes actuator dynamics for zero-shot sim-to-real transfer (arXiv 2607.02205 source); VLAFlow proposes a unified training framework for cross-comparing different robot pretraining paradigms (arXiv 2607.01586 source); Neuro-Symbolic Safety Guidance uses constrained flow matching to give VLA predictive obstacle avoidance (arXiv 2607.01378 source); Controllable Sim Agents generates traffic simulation agents controllable along interpretable axes via behavioral latent variables (arXiv 2607.02496 source).
Open Source · Tools · Benchmarks
· Embodied.cpp: A portable embodied-model inference runtime for heterogeneous robots, unifying deployment of VLA and world-action models (WAM), providing modular multi-rate execution, latency-first fused inference, and extensible operators/IO — easing the fragmentation of "each model with its own Python stack plus robot-side glue code" (arXiv 2607.02501 source).
· CommonRoad-Game: A lightweight human-in-the-loop autonomous driving simulation framework, tightly coupled with the CommonRoad platform, specifically built for systematically testing motion planners in interactive scenarios involving humans and analyzing human driving behavior (arXiv 2607.01382 source).
· DL-VINS-Factory: A modular framework unifying learned visual front-ends (ALIKED, SuperPoint, XFeat, etc.) with LK optical flow or LightGlue matching for visual-inertial SLAM, enabling systematic evaluation of the practical value of deep features in tightly-coupled VI-SLAM (arXiv 2607.01757 source).
2. Funding & Deals
XBOT | Series A + Series B | Hundreds of Millions of Yuan Combined · embodied
General-purpose food-service embodied robotics company XBOT closed two consecutive funding rounds: a ¥200 million Series A funded by Hong Kong's Jiankun Capital (GPTX), and a ¥300–500 million Series B with participation from multiple government funds, USD funds, and industry partners. The company was founded by former Xiaomi executive Tang Mu, and on the same day launched an in-house coffee brand to validate embodied commercialization through "a single cup of coffee." Funds will go toward R&D, market expansion, and team building.Source: Tech Capital Circle source (WeChat, CN), Robotics Outlook source (WeChat, CN)
Quanzhibo | Series A+++ | Led by GL Ventures · hardware
Wuxi-based Quanzhibo (a maker of integrated robotic joint modules) closed a Series A+++ round led by GL Ventures (Hillhouse's venture arm), with Zhiyuan Robotics (Chinese humanoid robot startup) and Lingxin Qiaoshou (Chinese dexterous-hand maker) joining as strategic industry investors. Both robot makers and top-tier capital are now betting directly on upstream joint modules, extending this cycle's trend of "capital flowing from full robots toward components like joint modules and dexterous hands."Source: PEDaily source, Shouchuang Holdings source (WeChat, CN)
Rushen Robotics | Pre-A Round | ¥100 Million · embodied
Shanghai-based Rushen Robotics (founded by a Tsinghua professor, focused on elderly-care scenarios for embodied AI) closed a ¥100 million Pre-A round, with investment from Qingsong Capital, Runze Technology, and Pinghu Zexin, to accelerate deployment of embodied AI in elderly-care facilities and home settings.Source: Robotics Outlook source (WeChat, CN)
Momenta (Shanghai Autonomous Driving Technology) | Hong Kong IPO | 414x Oversubscribed · autonomy
Autonomous driving solution provider Momenta's Hong Kong IPO public offering was oversubscribed roughly 414 times, adding another hot listing to the recent wave of Hong Kong IPOs. Investor enthusiasm for advanced autonomous-driving and embodied-AI names mirrors the rally in China's A-share robotics sector.Source: South China Morning Post source
Lishang LISSOME | Series A | Tens of Millions of Yuan · embodied
AI kitchen robot brand Lishang closed a tens-of-millions-of-yuan Series A round led by Sequoia China and Brizan Ventures, targeting embodied AI deployment in the high-frequency home cooking scenario.Source: Embodied Universe source (WeChat, CN)
3. Commercialization & Deployment
Kodiak Completes Autonomous Trucking Program in Ohio · autonomy
Kodiak announced the completion of its autonomous trucking program in Ohio, marking another real-world milestone for driverless long-haul freight; the autonomous trucking sector has recently been making steady progress on "mileage/program delivery," not just demos.Source: Investing.com source
Tesla Robotaxi Expands to Florida, Its Third Operating State · autonomy
Tesla carved out a small Robotaxi operating zone in Miami, making Florida its third state for autonomous-driving deployment. But media reports also note that Tesla is still struggling to scale up in Texas — a clear gap remains between expanding to new states and actually scaling capacity.Source: Electrek source
Waymo Takes First Step into Southern Europe · autonomy
Waymo has taken its first step into southern Europe, continuing to accelerate its overseas/international expansion. Following adjustments to its Phoenix partnership, Waymo's focus is increasingly shifting toward overseas/international cities and new partnerships.Source: eletric-vehicles.com source
UBTech's "Kissonic" Industrial Humanoid Debuts With 4,000 Pre-Orders · humanoid
UBTech (Chinese humanoid robotics maker) launched its industrial-native humanoid robot "Kissonic," accompanied by an industrial embodied model called "FabriX." The company claims it secured 4,000 units in orders at launch. The figure is a vendor-disclosed order count; actual installations and delivery pace remain to be verified.⚠️ Vendor-disclosed figuresSource: Xingfu Feidong source (WeChat, CN)
Zhejiang Humanoid "NAVIAI" Claims Scaled Commercial Use Across Four Sectors · humanoid
Ningbo-based Zhejiang Humanoid says its fully self-developed "NAVIAI" series has already achieved scaled commercial use this year across four sectors: industrial manufacturing, commercial services, education/training, and data collection. This is a company self-report; specific installation numbers and revenue have not been independently disclosed.⚠️ Vendor-disclosed figuresSource: Ningbo Cyberspace Affairs source (WeChat, CN)
4. Industry Developments
South Korean Conglomerates Collectively Bet on Physical AI: Samsung, Hyundai Pledge ~₩312 Trillion in Southern Investments · industrial
South Korean giants Samsung, SK, and the Hyundai Motor Group unveiled industrial blueprints for the Yeongnam region on the same day, with combined investments of roughly ₩312 trillion (about $195 billion), covering AI data centers, robotics factories, and launch vehicles. Samsung alone will invest ₩60 trillion to build a physical AI manufacturing cluster in Yeongnam, plus ₩19 trillion for a robotics factory in Gumi; Hyundai announced roughly $27.3 billion in investment in the southeast to develop mobility and physical AI. Following Japan's "sovereign AI" ¥1 trillion push and China's embodied-AI surge, South Korea is now formally joining the physical AI race with state-level capital and manufacturing bases.Source: Reuters source, The Korea Times source, Chosunbiz source
Estun Plans Full Acquisition of Estun Cobot to Expand into Embodied AI · industrial
Chinese industrial robotics leader Estun Automation disclosed it is planning a full acquisition of its collaborative-robot subsidiary Estun Cobot, widely read by the market as a signal of consolidation into embodied AI, with the stock hitting the daily limit-up 3 times in 4 days. Industrial robot makers expanding into collaborative/embodied businesses is another instance in this cycle's trend of full-robot manufacturers filling gaps in their portfolios.Source: JRJ source
Unitree IPO Approval Ignites China's A-Share Robotics Sector, Over 40 Stocks Hit Limit-Up · humanoid
Following yesterday's approval of Unitree Robotics' IPO registration on the STAR Market (already reported), China's robotics industry index rose nearly 8% intraday to lead the market today, with Green Harmonic Drive and more than 40 other stocks hitting limit-up, and robotics ETFs saw large net inflows. Multiple brokerages view this as a "catalyst and industry validation" for the sector, though some analysts warn of divergence risk between "the tail end of thematic speculation" and "the early stage of earnings validation."Source: Cailianshe source, Zhongjin Online source
Boston Dynamics CEO: Without a National Strategy, U.S. Robotics Risks Repeating Semiconductors' Fate · humanoid
As China, Japan, and South Korea successively elevate physical AI to a national strategy, Boston Dynamics' CEO publicly called for the United States to adopt a national-level robotics strategy, warning that without one, U.S. leadership in robotics could suffer the same fate as semiconductors. This is an executive statement, but it stands in sharp contrast to South Korea's ₩312 trillion investment announced the same day, highlighting an emerging "industrial policy race" among nations.⚠️ Executive statementSource: finance.biggo.com source
CloudMinds Unveils Single-Arm Collaborative Embodied Robot · embodied
CloudMinds (Chinese robotics maker) launched its first single-arm collaborative embodied robot, positioned around fusing world models and VLA as complementary approaches (rather than a either/or choice), targeting hotel and commercial service scenarios. The product integrates "predicting the future + semantic understanding + action" into a single collaborative platform.Source: Ikanchai source
AgiBot Unveils "Event-Level Predictive" World Model WALL-WM · world-model
AgiBot Robotics (Chinese embodied AI startup) says it has released the "world's first event-level predictive" embodied world model, WALL-WM, arguing for moving beyond "learning pixel changes frame by frame" toward modeling task-semantic events instead. The "world's first" claim is a vendor self-description; independent comparisons and third-party reproduction are not yet available.⚠️ Vendor-disclosed claimSource: Dolphin Digital Intelligence Lab source (WeChat, CN)
Guangzhou and Shanghai Roll Out Embodied AI Policies and Real-World Pilot Programs · industrial
Local governments continue to ramp up support: Guangzhou issued "Measures for Promoting the High-Quality Development of the Embodied Intelligent Robotics Industry," covering data collection standards, intelligent computing, and open scenarios; Shanghai's Municipal Commission of Economy and Information Technology launched a 2026 real-world pilot program for humanoid robots and embodied AI, targeting the identification of more than 100 high-value application scenarios by year-end and deployment capacity at the scale of tens of thousands of units. These include forward-looking targets, but represent genuine institutional efforts driving the industry forward. Separately, China's Ministry of Human Resources and Social Security plans to add 12 new occupations, including "embodied AI robot application technician."Source: Guangzhou Industry and Information Technology Bureau source (WeChat, CN), KCSJ Daily source (WeChat, CN), Jiemian News source
Tesla's Optimus Production Line Transition; VP Says Mass Production Begins by Year-End · humanoid
Musk posted a photo of the Optimus factory, with reports indicating the Fremont production line is transitioning from Model S/X toward Optimus; a Tesla vice president said humanoid robot mass production would begin by year-end, targeting an annual output of one million units. The line transition is an observable fact, while the "one million units" figure is a target/plan; the mass-production timeline remains to be delivered.⚠️ Planned/target figuresSource: Dasheng Liaowang source (WeChat, CN), chinanews.com.cn source
Hardware · Supply Chain
· Dexterous hand motors: Multiple technical reviews note that dexterous hands are essentially "multi-axis micro motion-control systems" (with roughly 20 joints operating simultaneously); motors are upgrading from coreless designs to miniature servos with built-in drivers to compress wiring harness and controller size, making this a key bottleneck for scaling dexterous-hand production.
· Dense dexterous-hand funding: Industry tallies indicate the dexterous-hand sector saw 3 funding deals totaling more than ¥800 million in the first week of July; dexterous hands previously accounted for less than 5% of humanoid full-robot funding, but this imbalance began correcting starting in Q2, with hardware focus shifting from "brain/torso" toward "hands."⚠️ Third-party tally
· Inspire-Robots: Cited as the global benchmark for dexterous-hand mass production, with annual shipments exceeding 10,000 units, covering core structures for miniature lead screws, servo cylinders, and tactile dexterous hands, along with supporting force-control SDK and tactile-data-collection software.
· Harmonic drive reducers: A single humanoid robot requires more than twenty precision reducers, accounting for about 15% of total unit cost; as 2026 mass production ramps up, the industry expects harmonic drive reducer output to grow at a compound annual rate exceeding 85%, opening a window for Chinese-made substitution.⚠️ Brokerage forecast
Top comments (0)