Today's Highlights
· Xi Jinping attends WAIC for the first time and sets the tone for AI moving into the physical world, giving embodied intelligence its highest-level policy endorsement to date
· AgiBot's Genie G2 Max partners with JD Logistics — the first time a humanoid robot has entered a real production warehouse setting
· Zoox issues a software recall for its robotaxi fleet after a thick-smoke misjudgment, covering 105 vehicles
· Xynova (Chinese dexterous-hand startup) raises a RMB 500M Series A+ led by Meituan, pulling in RMB 1.5B in 7 months to become a new unicorn in dexterous hands
· OFweek: mainland China's humanoid robot production topped 30,000 units in H1, with shipments of about 23,000 units
I. Research Progress
RoboTTT: Stretching Robot Policy "Context" by Three Orders of Magnitude · vla
Robotic foundation models have previously only been able to consume single-step or short-history visual-action context. RoboTTT is the first to extend this window to 8K timesteps — three orders of magnitude longer than existing policies — without slowing down inference. The approach embeds test-time training (TTT) inside a VLA: a set of "fast weights" continuously undergo gradient updates during both training and inference, compressing long history into weight space for retrieval as needed. Overall success rate on real-robot manipulation improved 87% over a single-step baseline, and the model independently completed a five-minute, ten-stage assembly task that all baselines failed at; the 8K context outperformed a 1K pretraining context by a further 62%. The authors propose "context length" as a new scaling axis for robotic foundation models.
Yunfan Jiang et al. (NVIDIA GEAR) · arXiv 2607.15275 source · HF↑11
A Scaling Recipe for Humanoid "Behavior Foundation Models" · locomotion
Whole-body humanoid control demands natural coordination, real-time responsiveness, and cross-environment generalization all at once. The team revisited the scaling laws for behavior foundation models (BFMs) and found three factors must be co-optimized: unifying diverse controls into motion tracking under a global coordinate frame, the ratio between online rollout count and reference motion diversity, and a scalable Humanoid Transformer architecture. In simulation and on real hardware, keypoint pose error (MPKPE) dropped 82% under the global mode compared with existing controllers.
Weishuai Zeng et al. · arXiv 2607.15163 source
DriftWorld: Pushing Robot World Models to 30+ FPS with "Drift" · world-model
Using world models for planning requires generating large volumes of rollouts quickly, but diffusion-based world models are expensive due to multi-step sampling at every call. DriftWorld instead learns an "action-conditioned drift" during training, so at inference a single forward pass generates future frames — averaging 17x faster than diffusion baselines and reaching 30+ FPS, while achieving SOTA decision-making performance with less inference time on benchmarks like Bridge-V2, RT-1, and Push-T. When used as an offline "policy scorer," its scores correlate with real success rates at up to 0.99.
Susie Lu et al. · arXiv 2607.15065 source · Analysis: 世模人 source (WeChat, CN)
When World-Action Models "Think Right but Act Wrong": A New Class of Adversarial Attack · world-model
World-action models (WAMs) couple action generation with future prediction, and the field has generally assumed that "verifying actions against imagined futures" makes them safer. BadWAM disproves this assumption's robustness: tiny visual perturbations can decouple the alignment between "imagination" and "execution" — an aggressive mode drives closed-loop success rate straight down from 96.5% to 43.1%, while a stealth mode leaves the model still "imagining" plausible-looking futures while executing misaligned actions, bypassing imagination-based safety checks entirely.
Qi Li et al. · arXiv 2607.15207 source · HF↑34
RxBrain: Merging Language Reasoning and "Visual Imagination" into One Embodied Cognitive Base Model · world-model
Unlike VLMs focused on scene understanding, or generative world models that only predict future frames, Tencent Hunyuan's RxBrain represents embodied plans as a single sequence: language provides an abstract skeleton of task decomposition, constraints, and decision logic, while visual imagination uses world-state prediction to ground each step in concrete physical states. It employs a unified Mixture-of-Transformers, enabling understanding and generation across language/image/video within a single model; when extended to continuous action, it achieves usable real-robot performance without large-scale action-data pretraining.
Haotian Liang et al. (Tencent Hunyuan) · arXiv 2607.14187 source
Reflex: Streaming Inference Lets Flow-Matching VLAs Output While Reasoning, Achieving Real-Time Control · vla
Flow-matching VLAs offer high precision for continuous control, but iterative denoising is inherently at odds with real-time robotics: injecting global timesteps breaks KV-cache, forcing a choice between slow recomputation and "mathematically incorrect cache reuse." Reflex restructures this process with streaming inference, compressing latency to real-time-executable levels while preserving continuous-control precision.
Yuanchun Guo, Bingyan Liu · arXiv 2607.14695 source
MIDAS Hand: A $3,000, 3D-Printed Open-Source Dexterous Hand · manipulation
The bottleneck in dexterous manipulation isn't only algorithmic — it's also the lack of hardware that balances human-like hand size, ease of manufacture and maintenance, tactile sensing, and controlled cost. MIDAS Hand delivers a 16-DoF (13 active) open-source hand with direct-drive low back-drive torque and 283 integrated triaxial tactile units, weighing 700 grams total, with a bill-of-materials cost under $3,000 and 3D-printed parts assemblable within three hours — released alongside design files, control/tactile APIs, simulation models, and a teleoperation pipeline.
Alvin Zhu et al. · arXiv 2607.14487 source
Other papers today: Open-AoE, an open first-person manipulation dataset and toolchain (arXiv 2607.14183 source); Wan-Streamer v0.3, a real-time full-duplex audio-video interaction model framing "video = world + event stream" (arXiv 2607.15038 source); Instant NuRec, feed-forward 3D Gaussian reconstruction accelerating autonomous-driving scene simulation, from NVIDIA (arXiv 2607.14203 source); Never Too Late for Force, using reactive force injection to accelerate VLA post-training (arXiv 2607.14236 source); KineFuse, kinematics-aware tactile fusion for in-hand pose tracking of occluded objects (arXiv 2607.14842 source); Active Real-World Factor-Based Evaluation, an active real-robot factorized evaluation for general robot policies (arXiv 2607.14439 source); Beyond Visual Grasping, a complex grasping benchmark spanning detection to execution (arXiv 2607.14341 source).
Open Source · Tools · Benchmarks
· HY-World 2.0: Tencent Hunyuan's open-sourced systematic multimodal world model, running a four-stage pipeline through understanding—synthesis—reconstruction, claimed to match closed-source commercial models source (WeChat, CN)
· MoWorld 3D: the first collaborative world model released by China's National AI Application Pilot Base (Embodied Intelligence), developed jointly with Huawei Cloud compute source (WeChat, CN)
· EmbodiedCLUE-World 2.0: an updated release of an embodied world-model evaluation benchmark, led by Peking University's PF-Cosmos, emphasizing physical laws and executable manipulation workflows source (WeChat, CN)
II. Financing & Deals
Xynova (Chinese dexterous-hand startup) | Series A+ | RMB 500M | Reaches Unicorn Status · embodied
Led by Meituan, with follow-on from NIO Capital, China Merchants Capital, and an unnamed major internet company, plus additional backing from existing strategic investor Xiaomi. Xynova makes dexterous hands for humanoid robots; founded in late 2024, it has completed 4 funding rounds totaling roughly RMB 1.5B in under two years. Its flagship product, Flex 2, uses a hybrid "tendon-driven + direct-drive" architecture with 23 degrees of freedom, a total hand weight under 400 grams, and fingertip force exceeding 20N. The company has built a 5,400-sqm factory in Hangzhou and plans to reach annual capacity of 10,000 dexterous hands and 200,000 micro linear actuators by the end of 2026 — the shift from "building one hand" to "reliably replicating ten thousand hands" is exactly the threshold for competing to become a Tier 1 supplier. Industrial capital including CATL, Xiaomi, JD, Li Auto, Meituan, and NIO is placing dense bets on this "last centimeter." Source: NE时代智能体 source (WeChat, CN)
MoSense (Chinese tactile-sensing startup) | Angel Round | Tens of Millions RMB · embodied
Jointly invested by Sequoia China, Hillhouse Ventures, and AgiBot. MoSense builds whole-body multimodal fusion tactile systems for robots; founded in May 2026, headquartered in Shanghai with an R&D center in Shenzhen's Qianhai. Tactile sensing is the "skin" that connects intelligent decision-making to the physical world — top-tier VCs joining forces with industry players continues this week's capital heat in the tactile/dexterous-hand sector. Source: 机器人前瞻 source (WeChat, CN)
Kunlunxing Robot | Three Rounds in 90 Days Since Registration | Tens of Billions RMB Cumulative · embodied ⚠️ single-source claim
Registered less than 90 days ago, the company has already landed three consecutive funding rounds, with Hillhouse Capital, HighLight Capital, and Chinaco Ventures each adding on. The company states its valuation now exceeds $1 billion, making it a unicorn — the fastest such run in the embodied-intelligence industry (pace and cumulative amount are company-disclosed figures). Source: 有材 source (WeChat, CN)
Humanoid (UK) | New Round | $150M | Unicorn · humanoid
UK humanoid robotics company Humanoid has closed a $150M round and reached unicorn status, adding another high-valuation player to Europe's humanoid race. Source: 智东西 source
X Square Robot (自变量机器人, Chinese embodied-AI startup) | Four Rounds Raised | Valuation Over $2.8B · embodied
Focused on embodied foundation models (Physical AI foundation model), the company has completed four consecutive funding rounds, with valuation surpassing $2.8B, placing it in the top-valuation tier of China's embodied-foundation-model leaders. Source: AOL.com source
BXI Robotics (半醒具身) | New Round | Tens of Millions RMB · humanoid
A humanoid robot ODM solutions provider; the investor is A-share listed company Suochen Technology, with funds earmarked for next-generation humanoid R&D and overseas market expansion — a listed company using an equity stake to enter the humanoid supply chain. Source: 机器人前线 source (WeChat, CN)
III. Commercialization & Deployment
AgiBot's Genie G2 Max Enters JD's Smart Wolf Warehouse: First Humanoid Robot in a Real Production Warehouse · humanoid
At WAIC on July 17, JD Logistics and AgiBot jointly unveiled the safety-rated, heavy-payload humanoid robot Genie G2 Max, announcing its deployment at JD Logistics' Smart Wolf Warehouse — officially described as the first time a humanoid robot has been deployed in a real warehouse production workflow rather than a trade-show demo. G2 Max is focused on inbound handling and depalletizing: it has human-like arm proportions and a heavy-load force-control system, can grip and move standard totes around the clock, and interfaces with upstream and downstream automation equipment. The Smart Wolf Warehouse is JD's self-developed "goods-to-person" system, with vertical racking over 10 meters tall and storage density about 4 times that of a traditional warehouse, already deployed at scale in Beijing, Wuhan, Chengdu, and the UK. AgiBot's Genie G2 had previously surpassed 15,000 units shipped (previously reported); this move advances "first at-scale deployment scenario" from a shipment figure to embedding in a real production line. JD Logistics says it plans to procure 3 million robots over the next five years. Source: JD Logistics (京东物流JDL) source (WeChat, CN)
UBTECH's Walker U1: Orders Top 13,000 Units Across All Channels, Deliveries Start in September · humanoid ⚠️ single-source claim
UBTECH's "full-size ultra-lifelike humanoid robot" Walker U1, which went on sale in June, had accumulated more than 13,361 units in orders across all channels as of June 30, with deliveries starting September 16. The full-body version is priced at RMB 159,800, with the series spanning RMB 119,800–990,000, and is marketed for companionship (explicitly stated as "not a boyfriend/girlfriend/partner," shipped with custom makeup, not eligible for seven-day no-reason returns). The order volume is notable, but figures come from company disclosure, and delivery and mass production remain to be proven out. Source: 维科网人形机器人 source (WeChat, CN) · 新识研究所 source (WeChat, CN)
Ant Group's Lingbo "Smart Pharmacy" Named a WAIC "Showcase Highlight" · embodied
Ant Group's Lingbo unit deployed a "one brain, many robots" architecture in a smart pharmacy scenario, with multiple robots coordinating on large-scale picking and sorting, earning recognition as a WAIC 2026 "showcase highlight"; described as China's first robotic pharmacy, it was also featured as a conference highlight — a real example of embodied intelligence deployed in a vertical commercial setting. Source: 上观新闻 source
Volkswagen's Moia to Pilot Robotaxi Service in Hamburg · autonomy
Volkswagen's mobility brand Moia has announced a robotaxi pilot launching in Hamburg, Germany, entering local European autonomous-mobility operations — another example of a legacy automaker running its own robotaxi service. Source: 디지털투데이 source
IV. Industry Developments
Xi Jinping Attends WAIC for the First Time, Frames AI as "Moving from the Digital World into the Physical World" · embodied
On July 17, the 2026 World Artificial Intelligence Conference and High-Level Meeting on Global AI Governance opened in Shanghai, with China's President Xi Jinping attending for the first time and delivering the keynote address, stating clearly that "artificial intelligence is moving from the digital world into the physical world." This marks the first time China's top leadership has linked AI's next phase to the "physical world" at the highest-level setting, and industry observers read it as the highest-level policy endorsement yet for embodied intelligence. Earlier signals had been building progressively: in March, the National Development and Reform Commission added "embodied intelligence" to its strategic emerging industries catalog, and in May the Ministry of Science and Technology released a draft Action Plan for Embodied Intelligence Innovation and Development. The speech, combined with the density of industry activity on WAIC's first day — leading companies racing toward capital markets, real factory settings opening up for large-scale validation — still leaves three hurdles unaddressed industry-wide: mass production, real paying use cases, and profitability models have yet to be broadly proven out. A policy green light solves "can it go public," not "can it make money." Source: 天瑞丰年 source (WeChat, CN)
Zoox Issues Software Recall After Thick-Smoke Misjudgment, Covering 105-Vehicle Fleet · autonomy
Amazon's Zoox issued a software update recall covering its entire 105-vehicle fleet after a robotaxi, in June, braked hard and became stuck at a fire scene shrouded in thick smoke with no cones set up to cordon it off — the update adds recognition and response capability for "emergency scenes obscured by thick smoke." No one was in the vehicle at the time and no one was injured; a remote operator ultimately reversed the vehicle out of the scene. This is Zoox's second recall, following a hard-braking issue in March 2025, and the timing is notable — the recall decision was made on July 7, one day before NHTSA's administrator sent a letter to the entire industry demanding robotaxis stop obstructing emergency responders; regulators explicitly characterized "failure to recognize emergency scenes" as a "functional insufficiency" rather than a rare edge case (following on NHTSA's earlier ultimatum, previously reported). Around the same time, San Francisco's mayor called for tighter regulation after a Waymo traffic incident — regulatory scrutiny on U.S. robotaxis is tightening. Source: TechCrunch source
OFweek: Mainland China's Humanoid Robot Production Topped 30,000 Units in H1, Shipments About 23,000 · humanoid ⚠️ institutional estimate
According to OFweek's "2026 Humanoid Robot Industry Insight Report," China produced over 30,000 units of humanoid robots and shipped roughly 23,000 units in H1 2026, with full-year output projected at around 50,000 units. Leading-producer tallies: AgiBot shipped its 15,000th unit in June, with a full-year estimate of about 20,000; Unitree produced roughly 11,000 units of a single model from January–May, with full-year shipments potentially reaching 15,000; UBTECH's Walker U1 has orders exceeding 13,000, deliveries starting in September; and LimX Dynamics' Foshan production line has annual capacity in the tens of thousands. Compared with roughly 16,000 units installed globally in 2025, the production curve is climbing rapidly — these are third-party report estimates. Source: 维科网人形机器人 source (WeChat, CN)
XPeng's MONA L03 Launches Globally in Munich, Starting at RMB 123,800, With Second-Gen VLA · autonomy
Following its second-generation VLA passing local certification in Germany (previously reported), XPeng's MONA L03 launched in Munich starting at RMB 123,800, with plans to reach 65 countries, marketed around "one autonomous-driving stack spanning China and Europe" — requiring only a small amount of local European post-training data to adapt. Chinese autonomous-driving exports are shifting from "simple replication overseas" toward pushing flagship-tier autonomous driving down into mainstream-priced vehicles. Source: 汽车之家 source
NIO Puts Its Automotive-Grade 5nm "Shenji" Chip into Robots, Launches Embodied-Intelligence Development Platform · hardware
NIO showcased at WAIC putting its automotive-grade 5nm chip, "Shenji," to use in robots, and launched an embodied-intelligence development platform, opening its in-vehicle compute and toolchain to robot manufacturers — repurposing its smart-vehicle compute and chip capabilities for the embodied-intelligence space. Source: 驱动之家 source
BrainCo Unveils "World's First" Brain-Controlled Robot Development Platform · adjacent ⚠️ demo claim
BrainCo debuted a brain-computer-interface-driven robot training platform at WAIC, claiming that 10 minutes is enough to achieve "mind control" of a robot (capability based on a live demo and vendor claims). Source: StreetInsider source
MIIT Standards Committee Holds Annual Plenary Session, Building Out Embodied Intelligence's "Rules Foundation" · adjacent
The 2026 annual plenary of the Ministry of Industry and Information Technology's Standardization Technical Committee for Humanoid Robots and Embodied Intelligence, alongside "Standards Week," was held in Shaoxing, Zhejiang, with NavInfo taking the lead as chair unit of the Data Standards Working Group; industry standards for dexterous hands are now open for public comment (previously reported). As the industry races toward mass production, it is simultaneously moving to fill in the foundational rules and data standards. Source: 砍柴网 source
Hardware & Supply Chain
· Dexterous Hand Price Cuts and Volume Ramp: GGII estimates that China's dexterous-hand shipments will surge from 19,200 units in 2025 to 70,200 units in 2026 (+260%), with downstream volume ramp forcing localization of core joint components like micro motors, driving a wave of financing in the micro-motor sector source (WeChat, CN)
· Joyson Electronic (均胜电子): unveiled core robot components including dexterous hands and hybrid solid-liquid batteries at WAIC, marking a Tier 1 auto-parts supplier entering the embodied-intelligence supply chain source
· Zhongda Leader (中大力德): holds roughly RMB 3.2B in orders from Unitree, supplying 63%–80% of its components, yet gross margin fell from 25.77% to 21.44% — "selling more, earning less," reflecting upstream price-war pressure source (WeChat, CN)
Top comments (0)