Today's Highlights
· Baidu's Apollo Go robotaxi begins road testing in London, public rides planned for 2027
· Morphi (Chinese embodied-AI startup) raises over RMB 1 billion in angel round, backed jointly by Alibaba and Tencent, valuation tops RMB 7 billion
· BeingBeyond's Being-H0.8 derives tactile sensing from 500,000 hours of human video
· Fei-Fei Li's team teams up with NVIDIA to extend robot context to 8K timesteps
· Shunde, Guangdong establishes China's first government bureau dedicated to embodied AI
I. Paper Progress
Being-H0.8: Deriving tactile sensing from human videos, without buying sensors · world-model
Tactile sensing has long been stuck at an impasse: the most scalable data source — first-person video — has no touch information, while data that does carry real tactile signals is hard to scale, since teleoperation is inefficient, exoskeletons are bulky, and gloves are noisy. BeingBeyond's (Chinese embodied-AI lab, joint Peking University/Renmin University team) solution bypasses sensors entirely: TactoHand uses roughly 30.1 million frames of hand-object interaction data to automatically generate two types of supervision — binary contact maps (touched or not) and continuous proximity fields (how close) — from the geometric distance and normal between hand mesh and object surface. It then applies this to label 500,000 hours of human video, cutting the cost of tactile supervision from "buy sensors and collect data" down to "run an algorithm." A companion module, TopoHand, uses a skeletal representation to unify human hands, dexterous hands, and parallel grippers into a single topology, allowing tactile experience learned from human videos to transfer to robotic hands. On the execution side, the system splits into a slow stream (planning overall motion rhythm) and a fast stream (fusing only the latest proprioception and touch for millisecond-level force adjustment). Demo tasks were chosen specifically to defeat pure vision — squeezing toothpaste, pinching grapes, feeling around inside an occluded bag. Worth noting: tactile sensing derived purely from geometry is still a visual byproduct at heart, and can distort under heavy occlusion, smooth materials, or subtle force — the team itself relies on real tactile sensor data as a fallback during the robot's post-training stage.
BeingBeyond (Chinese embodied-AI lab affiliated with Peking University/Renmin University, corresponding author Lu Zongqing) · Technical report, paper not yet public · Coverage: 具身智能之心 source (WeChat, CN), 具身纪元 source (WeChat, CN)
RoboTTT: Stretching robot "memory" from a few steps to 8K · vla
Current robot foundation models largely operate on a single step or very short history, even though context length has long been a core scaling dimension for large language models. Fei-Fei Li's team (Stanford) and NVIDIA's Jim Fan bring test-time training (TTT) into VLA: recurrent state is carried by "fast weights," updated in real time via gradient descent during both training and inference, compressing historical trajectories into weight space. Combined with sequential action enforcement and truncated backpropagation through time, memory usage grows only with segment length, allowing training over arbitrary context lengths within a fixed budget. Visuomotor context is thereby extended to 8K timesteps, three orders of magnitude beyond current state-of-the-art policies, without added inference latency. On long-horizon dual-arm assembly tasks, average completion score reached 79%, 87% higher than a single-step-context baseline; on a 5-minute, 10-stage gear assembly task, it was the only method to complete the task fully — all baselines failed. Performance kept rising as context scaled from 128 to 8K, with no sign of saturation.
Jim Fan (NVIDIA), Fei-Fei Li's team (Stanford), et al. · Coverage: 智东西 source
Data Pyramid: How should multiple embodied data sources actually be mixed · benchmark
Multimodal foundation models learn to see and speak by ingesting the entire internet — embodied agents have no such shortcut, since they need paired observation, physical-state, and action data. This paper arranges multiple sources — simulation, human video, teleoperated real-robot data — into a pyramid ranked by "distance from real interaction," and systematically discusses what each layer can provide, its cost, and how to mix them — a rare piece of structured thinking at a time when everyone is racing to scale up raw data volume. HF upvotes 29↑.
Yifan Ye, Yankai Fu et al. · arXiv 2607.24744 source
The curse of precision: high-precision assembly has its own data scaling law · manipulation
Scaling-law research on imitation learning has mostly focused on open-world generalization, while the relationship between data volume and precision in "closed-world" tasks like robotic assembly has barely been systematically studied. This paper derives a new scaling relationship, turning "how much data is needed to reach a given precision" from engineering intuition into a quantifiable metric — for teams doing production-line-grade precision work, this is more useful than yet another generalization leaderboard.
Cuijie Xu, Yuanfan Xu et al. · arXiv 2607.23108 source
Stress-testing LLM agents in a real chemistry lab, 4,608 trials · benchmark
AI evaluation has largely stayed focused on knowledge, reasoning, and generative planning, but scientific research genuinely demands reliable physical execution and evidence-driven plan revision. This work uses a robotic chemistry lab as a physical-world testbed, exposing 45 modular workstations as machine-readable skills and running 4,608 trials, turning "physical executability" and "evidence-driven replanning" into measurable quantities. The results point to a real gap in deployment readiness, and also offer a diagnostic framework for closed-loop improvement.
Lulu Guo, Yingkai Sun et al. · arXiv 2607.23045 source
FeelWorld: World models imagine futures that look plausible but violate physics · world-model
Purely visual world models excel at predicting appearance changes, but whether contact occurs, how force is distributed, and whether an object is slipping often aren't immediately visible in RGB frames — so the futures such models imagine end up "visually convincing but physically impossible." This work incorporates tactile state into the world model, performing hierarchical contact prediction and planning so the model can judge whether contact will actually occur before acting. Bringing touch into world models is a theme recurring across this week's papers, approaching the same problem from a different direction than Being-H0.8.
Wenxuan Ma, Chaofan Zhang et al. · arXiv 2607.24267 source · Coverage: ICCV source (WeChat, CN)
Picking the "optimal" trajectory from a world model can backfire · world-model
Sampling-based controllers using latent world models follow a conventional rule: assign a terminal cost to each candidate action sequence, take the one with the lowest cost, execute its first segment, then replan. This paper shows the rule fails — even when the terminal cost perfectly and accurately reflects the true task objective. The authors propose ASAR, which instead estimates density from standardized early-action prefixes, reconstructs the full sequence from a "neighborhood" set, and uses the minimum-cost sequence only as a lightweight anchor. This is a fundamental challenge to how world models are used downstream — more worth reading than another marginal SOTA bump.
Liangyu Li, Qingwen Liu et al. · arXiv 2607.23602 source
Can GPT-5.1, which has never touched a robot, serve as the brain of a mobile robot? · world-model
An exploratory study: using GPT-5.1 directly as the high-level controller for a physical mobile robot, with the model having no embodied experience whatsoever — no simulation training, no exposure to sensorimotor data. The question the results point to is more interesting than the results themselves — just how much usable world knowledge general multimodal models "pick up along the way" from pure language pretraining alone.
Roberto Spinelli, Thiago C. Martins · arXiv 2607.23899 source
Other papers today: LeapBot-WA uses predictive latent-space alignment for a world-anchored action model, sidestepping the compute waste of pixel-level video generation (arXiv 2607.23969 source); WorldDiT uses a unified diffusion transformer to couple action generation with visual world modeling, without relying on a large pretrained VLM backbone (arXiv 2607.23909 source); PRISM explicitly represents polynomial interactions between observable physical quantities as a representation choice alongside MLP policies (arXiv 2607.23473 source); a renormalization approach to sim2real uses resolution-dependent effective parameters to absorb details omitted by simulation, rather than simply building ever-more-precise simulators (arXiv 2607.24079 source); ArmnetBench v0.1 uses a batch of low-cost SO-101 units for parallel real-robot evaluation, easing the hardware bottleneck in policy evaluation (arXiv 2607.24481 source); KAI improves data efficiency for articulated-object manipulation using a kinematics-aware intermediate representation (arXiv 2607.24493 source); Instance-Oriented Memory lets a robot explore an object with hidden state (e.g., a locked microwave) only once (arXiv 2607.23702 source); LabRobFail releases a robot failure-analysis benchmark for autonomous chemistry labs (arXiv 2607.23704 source); pose-dependent artifacts in tactile gloves can be mitigated algorithmically using hand pose information, without hardware changes (arXiv 2607.22964 source); arXiv versions of N₀-VTLA and N₀-TWAM posted the same day (arXiv 2607.23782 source, 2607.23783; the models themselves were previously reported).
Open Source · Tools · Benchmarks
· TriWorldBench: a joint initiative from Peking University, Tsinghua, Beihang, Shanghai Jiao Tong University, and USTC, this is the first benchmark for robot world models covering three viewpoints (head + left/right wrists) — with 500 tri-view synchronized episodes and 50 manipulation tasks, aggregating 19 signals across six dimensions; the core test is whether the three video streams describe a consistent world (arm state, grasp phase, and object state not contradicting each other) rather than each merely "looking realistic" on its own — open registration for teams globally source (WeChat, CN)
· τ0-VLA: the previously reported "imagine-then-act" long-horizon VLA is now officially open-sourced source (WeChat, CN)
· Harness VLA: open-sourced by a Tsinghua team, this wraps existing VLAs into a VLA_ACT primitive dedicated to fine-grained contact manipulation without retraining the underlying policy, adding a scheduling layer on top source (WeChat, CN)
· Humanoid Motion Control Knowledge Base: an open-source collection spanning everything from motion retargeting to world models, aimed at readers from beginners to technical leads source (WeChat, CN)
II. Funding & Deals
Morphi (Chinese embodied-AI startup) | Angel Round Series | Over RMB 1 billion | Post-money valuation over RMB 7 billion · humanoid
Jointly invested by Alibaba and Tencent, with participation from BlueRun Ventures, Legend Capital, Huaye Tiancheng, Banyan Capital, Casstar, and Source Code Capital, with Gao Hu Capital serving as exclusive financial advisor. It's common for the two internet giants to co-invest at mid-to-late stage, but it's rare for them to jointly back the same hard-tech startup at the angel round. The company registered in August 2025, meaning it was less than a year old at the time of announcement. CTO Huang Qingqiu is a "Huawei Genius Youth" recruit and former head of Huawei's automotive BU intelligent-driving AI division, leading development of multiple generations of ADS algorithms. Its technical approach is full-stack self-developed hardware and software (embodied foundation model, world model, VLA, and bipedal platform), with a commercialization-first deployment strategy prioritizing hotel services and last-mile delivery. Source: 湾财ND source (WeChat, CN)
Delta Intelligence (Chinese world-model startup) | Angel++ Round | Nearly RMB 500 million · world-model
Founded in January 2026, this marks the company's sixth funding round in six months. Previous investors include the Beijing Institute of General Artificial Intelligence and humanoid manufacturers Zhiyuan (Chinese humanoid startup), Leju, and Galbot, as well as Hillhouse Ventures, Yuanhe Holding, and Legend Capital. All three founders hold PhDs from UCLA, with backgrounds spanning Google Robotics, NVIDIA Research, DeepMind, and Meta. Its technical bet is on "native 3D": the team argues that mainstream embodied models still ground spatial cognition in 2D visual representations, meaning a robot standing in front of a door can't judge handle distance or how much clearance it needs to open it, with errors accumulating over long-horizon tasks. Its approach instead directly processes 3D representations such as point clouds and Gaussian splatting, paired with a proprietary full-body panoramic data-capture rig (synchronously capturing skeletal trajectories of hands, feet, waist, and shoulders), and a three-layer "brain + cerebellum + force-position hybrid" architecture for millisecond-level control. Source: 36Kr source (WeChat, CN)
Greyparrot (UK) | Series B | $27 million · industrial
Led by Om Elite Assets Investment. The product, Greyparrot Analyzer, is an AI camera mounted above sorting-line conveyor belts, capable of identifying 111 types of waste and having logged over 1 trillion detections to date. Its angle is economic rather than environmental — pulling aluminum, PET, and HDPE out of waste streams to process into clean raw materials for resale, with the company claiming up to 40x return on investment for recycling facilities. This is a typical example of physical AI applied to legacy heavy-asset production lines — "add sensing, don't replace equipment."Source: SiliconANGLE source
Honeywell Warehouse Automation business | Acquisition completed | Amount undisclosed · industrial
American Industrial Partners has completed its acquisition of Honeywell's Warehouse and Workflow Solutions business, merging Intelligrated, Trew, and Transnorm into a single warehouse automation platform. Intelligrated provides sorting, conveyance, palletizing, robotics, and warehouse software, with over 3,300 employees; Trew was already an AIP portfolio company, while Transnorm will continue operating independently. Alfred Rebello will serve as CEO of the combined entity.Source: citybiz source
DexForce Technology (Shenzhen) | Planning Hong Kong IPO | Possibly raising hundreds of millions of dollars · embodied
According to Bloomberg, citing people familiar with the matter, this Shenzhen-based robotics and AI company founded in 2021 has held preliminary discussions about a listing and potential advisors; its products cover industrial automation, home appliances, and logistics/warehousing. Its previous Series B raised RMB 1 billion at a post-money valuation of over RMB 10 billion, with investors including Shenzhen Capital Group and ICBC Capital.⚠️ Unconfirmed by the companySource: The Standard source
Yunmu Zhizao | Pre-A Round | Tens of millions of RMB · humanoid
Backed by Huaqiang Ventures. The team has an aerospace background, and is pursuing three product lines simultaneously: highly realistic IP-based humanoids, industrial embodied humanoids, and specialized embodied humanoids.Source: Huaqiang Ventures source (WeChat, CN)
III. Commercialization & Deployment
Baidu's Apollo Go launches robotaxi road testing in London, opening to the public in 2027 · autonomy
Baidu announced Tuesday that Apollo Go has begun road testing in the Brent area of northwest London, using its sixth-generation autonomous vehicle, the RT6, with the test zone covering a mix of urban and suburban road conditions and a safety driver still on board. Its partner is Freenow, Lyft's European mobility platform, which will handle local operations; both parties have held extensive discussions with Transport for London (TfL) and the UK's Centre for Connected and Autonomous Vehicles (CCAV), with plans to carry passengers by 2027 pending regulatory approval. This follows progress in Hong Kong five days earlier — on July 23, Apollo Go obtained the Hong Kong Transport Department's first fully driverless testing permit, began testing on Airport Island on the 27th, marking the world's first public-road, fully driverless trial in a right-hand-drive market; London, also right-hand-drive, serves as a parallel technical validation site. In scale, Apollo Go had completed over 22 million rides cumulatively as of April, with 3.2 million fully driverless orders in Q1 alone — up over 120% year-on-year; as of May it covered 27 cities globally, with cumulative autonomous driving mileage exceeding 330 million kilometers. A March survey by Freenow found that 48% of Londoners have never ridden in an autonomous vehicle, but 58% said they'd be willing to try one, with the figure approaching two-thirds among respondents under 35.Source: CnEVPost source, TechCrunch source
SAIC launches purpose-built robotaxi production vehicle program, to market by 2027 · autonomy
This is Shanghai's first mass-production vehicle model developed from the ground up to L4 driverless standards for the driver's seat, with dedicated design for sensor layout, computing platform redundancy, and body safety structure — moving away from the "retrofit a production car" approach. The project splits work three ways among an automaker, an autonomous-driving company, and a mobility platform: SAIC handles mass production, Momenta provides the L4 system, and Xiangdao Mobility leads product definition and subsequent operations. Xiangdao's robotaxi service has already logged over 350,000 user rides, 4 million kilometers of safe driving, and nearly 50,000 open operating points. Also released at the same event were an updated version of the "Shanghai Intelligent Connected Vehicle Testing and Demonstration Implementation Measures" and two local standards.Source: Shanghai Automobile Industry Association source (WeChat, CN)
South Korea's Daeduck Electronics puts humanoid robots on a PCB production line, a first · humanoid
Daeduck Electronics will introduce AeiROBOT's mobile humanoid ALICE M1 for a proof of concept, deploying 2 units in phases, with the first round of demonstrations to be completed by year-end. The ALICE M1 features dual seven-degree-of-freedom arms, a wheeled autonomous mobile base, and an adjustable waist (height varying between 1.3–1.7 meters); initially it will handle only repetitive manual work like moving and organizing PCBs, with the company deciding whether to expand deployment based on task success rate, walking paths, and cycle time — first results are expected in about six months. The rationale for choosing a humanoid is specific: PCB size, thickness, and layer count vary by customer and product, making dedicated automation equipment cost-prohibitive, while a camera-and-AI-equipped humanoid can learn to recognize different board types. This is South Korea's first humanoid deployment at a PCB plant; previous global humanoid pilots have concentrated mainly in the automotive industry. AeiROBOT plans to supply around 20 units this year for field data collection, moving to commercial scale from 2027.Source: The Elec source
Boston Dynamics' Atlas to ship to its first external customer next year · humanoid
According to industry sources, Boston Dynamics disclosed at a technical briefing on the 22nd that it will deliver Atlas to a global enterprise for a first pilot next year — marking Atlas's first deployment outside the Hyundai Motor Group. The goal is to put the robot, which has been trained on data from its parent company's manufacturing operations, onto an unfamiliar production line to gather new data. The team said multiple Atlas units at the same customer site will form a form of "collective intelligence": once one unit learns a new manipulation skill, that capability automatically propagates to every machine at the site. Another focus area is non-verbal communication — when a worker enters a 2-meter safety zone, the robot turns and flashes a light as a warning, and if they get closer still, it sets its load down on the ground first. When pausing to process information, it performs a head-scratching gesture, signaling to the worker that it hasn't malfunctioned.Source: Seoul Economic Daily source
Komatsu's 1,000th autonomous haul truck goes into operation, at Barrick's Nevada gold mine · industrial
Komatsu commissioned its 1,000th autonomous haul truck on April 21, becoming the first equipment manufacturer in history to reach that number — it's also the manufacturer that first commercialized an autonomous haulage system, back in 2008. The truck is a 930E-5AT electric-drive haul truck with a 290-metric-ton payload, also the most widely deployed model in the FrontRunner fleet (over 500 units in service). As of mid-2024, roughly 2,080 units of autonomous haul trucks were operating in open-pit mines worldwide, with Komatsu accounting for nearly half; Australia led with 927 units, followed by China, Canada, and Chile. Since FrontRunner's commercial launch, customers have collectively hauled over 11.5 billion metric tons of material.Source: ecoportal source
IV. Industry Developments
Shunde, Guangdong establishes China's first government bureau dedicated to embodied AI · industrial
On July 28, the Shunde District Embodied AI Development Bureau in Foshan was officially unveiled — described as the first dedicated government agency in China tasked with coordinating the entire embodied-AI industry chain, headed by Director Hai Yubao. Shunde wasn't chosen by accident — it's China's largest industrial-robot production base, home to a national-level specialized robotics small-and-medium-enterprise cluster designated by the Ministry of Industry and Information Technology, with the district's robotics industry generating nearly RMB 40 billion in output value in 2025. The bureau operates through four coordination mechanisms: an inter-departmental joint conference for overall coordination, an expert think tank for technical vetting, an industry alliance to link enterprises, and an innovation center to empower the platform. District Party Secretary Chen Xinwen said the bureau's purpose is to "resolve the pain point of technology being disconnected from industry." Local governments elsewhere have previously mainly set up funds, industrial parks, or application scenarios; establishing embodied AI as a standalone permanent government bureau is a new administrative form — whether it can genuinely link policy, technology, application scenarios, and talent remains to be seen.Source: 长安街知事 source (WeChat, CN)
LG Innotek teams up with TDK to build robots' "eyes" and "skin" · hardware
The two companies have entered a strategic partnership in physical AI: LG Innotek contributes optics and ultra-precision module design, while TDK brings its two decades of sensor expertise (inertial, orientation, temperature, pressure, position). The first step is integrating TDK's IMU into LG Innotek's vision-sensing module, using motion, position, and acoustic data to compensate for the shortcomings of pure image sensors — achieving motion-blur compensation and image stabilization, and avoiding visual blind spots via high-frequency tracking and dead reckoning; this module is planned for launch in 2027. The direct benefit for robot manufacturers is no longer having to source and connect multiple sensors individually. The second step, also targeted for 2027, is a tactile-sensing module to be mounted on robot hands, arms, and torsos, functioning similarly to human skin. The alliance between these two Japanese and Korean component giants points to a fight for influence over this modular-supply layer of robot perception.Source: PR Newswire source
MIIT: robotics-sector revenue tops RMB 90 billion for Jan–May, up 26.9% year-on-year · industrial
Data from China's Ministry of Industry and Information Technology shows that revenue at China's above-scale robotics enterprises exceeded RMB 90 billion from January to May this year, up 26.9% year-on-year, with average annual growth exceeding 20% over the past 5 years. At the same day's State Council Information Office briefing, He Shaojun, deputy director of the Ministry of Commerce's Foreign Trade Department, cited customs statistics showing industrial robot exports grew 18.6% year-on-year in the first half, while 3D printer exports grew 109.3%; Deputy Commerce Minister Yan Dong, responding to "China Shock 2.0" narratives, said they didn't reflect the facts. Listed companies' half-year reports and earnings forecasts are being released in batches, with strong showings across the robotics sector spanning the chain from core components and system integration to motion control and AI computing hardware.Source: 观点网 source, 新浪财经 source
Unitree: shipments to at least double this year, 70% of humanoid revenue from research and education · humanoid
According to Nikkei Asia, Unitree's (Chinese humanoid and quadruped robot maker) director for Asia-Pacific and other regions, Owen Chen, said at the Nikkei Global Digital Summit 2026 that total shipments in 2026 are expected to at least double compared to last year's over 6,500 units, and could even nearly triple; to meet orders, annual production capacity may expand to 30,000 units, with plans to use IPO proceeds to build a new factory with roughly 20,000-unit annual capacity. All current production capacity is in China, and while overseas assembly could be considered if demand keeps growing, he stressed "there are no concrete overseas production plans at this time." More notable is the revenue breakdown: over 70% of humanoid-robot-related revenue comes from research and education, while industrial applications account for only 9% — Owen Chen defined the former as the "developer market," stating outright that a robot operating system and applications have yet to emerge. This also serves as a direct footnote to the current humanoid mass-production narrative.⚠️ Executive statementSource: 智财社 source (WeChat, CN)
Zhiyuan's WITA-Omni Preview claims top spot on the DailyOmni multimodal understanding leaderboard · world-model
Zhiyuan (Chinese embodied-AI startup) released a preview of its embodied foundation model WITA-Omni, claiming first place on the DailyOmni audiovisual reasoning benchmark, ahead of Gemini and Doubao, emphasizing a shift from "being able to talk" to "understanding interaction" through joint audio-visual comprehension.⚠️ Vendor claimSelf-reported leaderboard results are a company claim rather than independently verified, and should not be treated as conclusive before third-party reproduction.Source: Zhiyuan official source
South Korean government pushes full speed on physical AI and robotics national programs, Hyundai Wia invests KRW 400 billion · industrial
The South Korean government announced it will fully advance a large-scale physical AI and robotics program; separately, Hyundai Wia announced a KRW 400 billion investment in its robotics business, expanding production facilities and logistics automation technology. Combined with today's LG Innotek–TDK sensing alliance and Daeduck Electronics' humanoid pilot at a PCB plant, South Korea's moves in physical AI are shifting from isolated trials toward simultaneous progress on policy, components, and applications.Source: newsworld source, Seoul Economic Daily source
Starship delivery robot sued over injury · adjacent
A woman has filed suit against Starship Technologies, alleging she was struck and seriously injured by one of its sidewalk delivery robots. Sidewalk delivery fleets have been expanding rapidly of late — just two days earlier, Uber Eats and Serve Robotics announced plans to deploy up to 2,000 units — pushing questions of liability over shared human-robot sidewalks into the spotlight.Source: Retail Technology Innovation Hub source
BYD confirms humanoid robot reveal in August · humanoid
BYD confirmed to reporters on July 28 that its first humanoid robot will formally debut in August, going into service at "Di Space" in Zhengzhou (previously reported). The news boosted A-share humanoid-robot-related stocks, with Jiezhong Technology and Mingxin Xuteng both hitting their daily limit up.Source: 财联社 source
Hardware · Supply Chain
· Frameless motors: one company reports that orders on hand for its frameless motors used in humanoid joint drives topped 1 million units in the first half of the year, up more than 9x year-on-year; another company says its new humanoid product received tens of thousands of orders less than a month after launch ⚠️ Vendor claim source
· Harmonic reducers: the price war for complete units, sparked by Unitree, is passing cost pressure upstream, forcing harmonic reducer suppliers out of their comfort zone — companies like Leaderdrive are accelerating product-line diversification source
· Humanoid-grade electrical steel: People's Daily reports that humanoid robots have entered the stage of industrial-scale mass production, and the high-grade non-oriented electrical steel required for their joint motors has become a new growth area for Chinese domestic materials source (WeChat, CN)
· Fully direct-drive dexterous hand: Daxian Shoushou unveiled a high-DOF, fully direct-drive dexterous hand at a WAIC forum, a design that avoids the wear and hysteresis issues of tendon-driven transmission — dexterous-hand lifespan has been repeatedly flagged this week as a bottleneck source (WeChat, CN)
Top comments (0)