Today's Highlights
· Morphi's angel round exceeds RMB 1 billion, valuation tops RMB 7 billion, with former Huawei autonomous driving lead Huang Qingqiu joining
· AgiBot's 15,000th embodied robot rolls off the line, annual capacity targeting 40,000-50,000 units
· Shengshu Technology raises RMB 3.4 billion in Series B+, China's largest single financing round for a world model
· Fudan's SRPO fine-tunes with just 200 RL steps, pushing VLA success rate to 99.2%
· UBTECH's RMB 990,000 companion robot gets only 2-4 hours of battery life; company says this is standard across the industry
I. Research Progress
MemoBench: A New Benchmark Exposes World Models' "Memory Blind Spots" · benchmark
The buzzy "world model" category has collectively been caught short by a new benchmark: MemoBench specifically tests whether video generation models have the "object permanence" capability humans possess by age three — can a model track an object that temporarily leaves the frame and keeps changing while occluded (ice melting, powder being poured), and correctly restore its "expected state" when it reappears. The team tested 10 mainstream world models on 360 ground-truth video segments, and none scored above 0.6 on "object reappearance" (out of 1.0), showing that even photorealistic output doesn't necessarily mean the model is truly "remembering and simulating" the world. This gives next-generation world models a quantifiable diagnostic yardstick.
Haoyu Chen et al. (Harvard, MIT, Google, etc.) · arXiv 2606.27537 source · via Jiqizhixin source (WeChat, CN)
SRPO: Making VLA "Self-Referential," 200 RL Steps Double Success Rate to 99.2% · vla
Reinforcement learning post-training for VLA has long been stuck with a binary success/failure signal, wasting large numbers of failed trajectories that actually contain partial progress. Fudan University's team led by Qiu Xipeng takes a different approach with SRPO: it uses successful trajectories generated within the current batch as "self-reference," measuring in the world model's latent space how far a failed trajectory progressed, and assigns dense rewards accordingly — no external demonstrations or manual reward engineering required. Starting from a supervised baseline of 48.9% on LIBERO, just 200 steps of RL fine-tuning pushed it to a new SOTA of 99.2%, with a 167% improvement on LIBERO-Plus.
Wang Siyin, Fei Senyu et al. (Fudan University, Shanghai Innovation Institute; corresponding author Qiu Xipeng) · CVPR 2026 · via Fanfan Reading PhD Diary source (WeChat, CN)
StableVLA: The Fragile Part Isn't the "Brain" — It's the "Translator" in Between · vla
As soon as robots enter the real world they face blur, glare, grime, and occlusion, and even state-of-the-art VLA models suffer sharp performance drops. Stardust Intelligence, together with Peking University and Tsinghua University, found the weak link isn't the vision encoder or the large language model — it's the Projector, the middle module that "translates" images for the language model. Drawing on information theory, they propose a lightweight IB-Adapter (information bottleneck adapter) that applies channel-wise "smart noise-reduction switches" to features, improving performance by an average of 35.2% across various visual disturbances. More notably, with just 500 million parameters (14x smaller backbone, and without Open X-Embodiment pretraining), it approaches the robustness of a 7-billion-parameter SOTA model. The work has been accepted to ICML 2026 and open-sourced.
Stardust Intelligence (Chinese embodied-AI startup) / Peking University / Tsinghua University · ICML 2026 · via Embodied Intelligence Research Lab source (WeChat, CN)
Do as I Do: First End-to-End Pipeline from Internet Videos to Real Dexterous-Hand Robots · manipulation
Huge volumes of human hand-manipulation videos already exist online; the hard part is turning noisy monocular RGB video into trajectories a dexterous hand can execute. UC Berkeley's Do as I Do is the first to complete this pipeline: it reconstructs 4D hand-object interaction, then retargets it to a 22-degree-of-freedom Sharpa Wave dexterous hand, generating 500 validated trajectories across 20 manipulation categories including scrubbing, spreading, whisking, and hammering, and deploying 10 real-world tasks at 50Hz on a bimanual platform. Retargeting success rate for noisy reference trajectories rose from 25% to 71%. The team also notes that only about one in twenty unfiltered internet videos is actually usable.
UC Berkeley · arXiv 2606.19333 source · via AI Era (Xin Zhi Yuan) source (WeChat, CN)
KEMO: Giving VLA an "Event Memory" That Only Remembers What Truly Matters · manipulation
In long-horizon manipulation, robots often get stuck because the current frame alone can't reveal which step of the task they're on. Rather than having the model remember all of history, KEMO (event-driven keyframe memory) automatically selects sparse keyframes triggered by interaction-induced state changes and stores them in a memory bank, then injects them into the current decision via gated cross-attention. It's a lightweight module that can plug directly into existing VLA policies, letting a robot remember progress markers such as "which cup it has covered so far" during multi-stage tasks.
Yihan Zeng et al. · arXiv 2606.23589 source · via Embodied Intelligence Heart source (WeChat, CN)
DriveTeach-VLA: Teaching Autonomous-Driving VLA "What to Look At and Where" · autonomy
Autonomous driving VLA models often receive abstract language instructions without knowing "where to look." Beihang University / Tsinghua University / Didi's DriveTeach-VLA feeds 2D text-guided points (readable geometric coordinates) directly into the planner, giving the model an alignable target location in image space, achieving PDMS of 90.4 on NAVSIM.
Beihang University / Tsinghua University / Didi · ECCV 2026 · via Vehicle Engineering Research source (WeChat, CN)
Open Source · Tools · Benchmarks
· ACE-Brain-0.5: Daxiang's open-source unified embodied foundation model, an 8B-parameter model handling perception/understanding/planning/action all in one; the company claims top global rankings on multiple metrics ⚠️ vendor-reported figures source (WeChat, CN)
· EVA-Client: Open-sourced by Beihang University, positioned to fill the "last mile" gap in embodied-AI open-source deployment/client tooling source (WeChat, CN)
II. Funding and Deals
Morphi (Chinese embodied-AI startup) | Angel Round | Over RMB 1 billion | Valuation over RMB 7 billion · embodied
Alibaba, Tencent, Lanchi Ventures, Legend Capital, Huaye Tiancheng, GaoRong Capital, Sino Star and others participated, setting a record for the largest publicly disclosed first-round financing in the Chinese embodied-AI sector for a company just six months old. CTO Huang Qingqiu, a former Huawei "Genius Youth" and head of Huawei's autonomous-driving AI unit, led Huawei's ADS1.0→4.0 autonomous-driving stack and is credited as the industry's first to mass-produce a one-stage end-to-end system at over a million units. The company is pursuing a "software-hardware integrated systems engineering" approach targeting general-purpose home robots, with its first service robot launching in July. Source: 36Kr source (WeChat, CN)
RobotEra | New Round | RMB 1 billion | Valuation over RMB 10 billion · humanoid
Led by Chengtong Fund, with state-owned capital from Jiangxi Guokong, Guoyuan Equity, Yufu Zhongxin, Hangzhou Capital and others joining, alongside CICC Renault, Ninetowns, and Hony Capital as co-investors. After raising roughly RMB 2.5 billion in a burst of financing over two months and nearly RMB 5 billion for the year to date, this Tsinghua-affiliated company founded by Chen Jianyu reportedly entered thousand-unit-scale delivery in Q2, with growth of roughly 300%. Source: Robot Qianzhan source (WeChat, CN)
Shengshu Technology | Series B+ | RMB 3.4 billion (approx. USD 500 million) · world-model
This sets a record for the largest single financing round for a general-purpose world model in China, riding the recent wave of capital pouring into world models (Goldman Sachs has said world models could become a new pillar of AI infrastructure, though some observers note the term still lacks a unified definition and early bubble signs). The company's unified world model for robot control, Motus, is claimed by the company to have an average multi-task success rate 35.1 percentage points higher than Pi0.5. Source: Zhidongxi source
Mouseshen Intelligence | New Round | RMB 300 million · embodied
Founded by Chen Yongchao, a scholar born after 1995 who turned down an offer from DeepMind, the company focuses on robot "brains," betting on a self-evolving large-model technical approach. Source: 36Kr source
Tripo AI | Series A3 | USD 150 million · world-model
Positioned as an AI-native 3D / general-purpose world model company, building production-grade infrastructure for 3D asset generation (gaming, film, industrial, AR/VR). This round drew joint backing from a major gaming company and an automaker. Source: Sina source
Xingfan Intelligence | New Round | Over RMB 300 million · embodied
Funds will be used to accelerate embodied-AI products and real-world deployment aimed at "bringing AI into the physical world." Source: PEdaily source
III. Commercialization and Deployment
AgiBot's 15,000th Embodied Robot Rolls Off the Line, Annual Capacity Targeting 40,000-50,000 Units · humanoid
On June 28, AgiBot's 15,000th general-purpose embodied robot, the Genie G2, rolled off the line in Shanghai — less than three months after the 10,000th unit, with mass-production pace still accelerating. Real production-line data backs its quality system: during a 6-day livestream at a Longqi Technology factory, it logged 64 hours of continuous operation, 64,828 cycles, and a 99.99% success rate; on SAIC-GM Buick Zhijing E7's battery mass-production line, it handles high-precision work with ±0.1mm positioning accuracy at about 2 seconds per unit. The company has built an "A-Chain" standardized supply-chain system, with over 20 listed companies confirmed as partners and flexible capacity of over 100,000 units/year; partner Yao Maoqing revealed this year's capacity is expected to reach 40,000-50,000 units. According to Omdia, AgiBot ranked first globally in 2025 shipments with 5,168 units (roughly 38.7% market share). While most players are still "struggling to reach a thousand units," this stands as one of the most solid mass-production milestones to date. Source: Yicai source (WeChat, CN)
Hikrobot's Full Range of Mobile Robots Surpasses 200,000 Units Off the Line · industrial
Hikrobot announced that its mobile robots across all product categories have surpassed 200,000 units off the production line — seen as an inflection-point signal for mobile robots reaching scaled deployment, with mass-production volume already at scale in logistics and warehousing scenarios. Source: Sina Finance source
Zhijian Power, 11 Months Old, Puts 100 Robots on Production Lines · industrial
Zhijian Power, founded just 11 months ago, has achieved the "fastest delivery of 100 units," integrating 100 robots into production-line cadence. Compared to a single working prototype, turning it into a batch of products with consistent performance and controlled cost is a much harder bar to clear — reaching 100 units on the line is a concrete step from demonstration to stable production. Source: Robot Qianzhan source (WeChat, CN)
McLaren Deploys Autonomous Robots at UK Construction Site via FieldAI · embodied
McLaren has partnered with embodied-AI company FieldAI to deploy autonomous mobile robots at a UK construction site, bringing general-purpose embodied capability into the unstructured, high-disturbance real-world setting of construction. Source: Construction News source
Multiple Waymos Run Over Fireworks in San Francisco on July 4th, One Catches Fire · autonomy
On July 4 in San Francisco, several Waymo robotaxis drove over fireworks debris on the road, with one catching fire; separately, reports indicate a Waymo vehicle ran out of battery amid holiday gridlock and had to be manually towed away. Once robotaxis hit real-world roads, the failures that surface are often not a single module bug but combinatorial failures across perception, prediction, and planning in rare scenarios — again putting safety redundancy and operational dispatch under scrutiny. Source: IT Home (Qudong Zhijia) source
IV. Industry Developments
UBTECH's RMB 990,000 Companion Robot Gets Only 2-4 Hours of Battery Life; Company Says "Standard Across the Industry" · humanoid
The Youshijie U1 bionic companion robot, priced at RMB 880,000-990,000, was reported to have only 2-4 hours of battery life. On July 6, UBTECH responded that "full-size humanoid robots generally get 2-4 hours of battery life across the industry," emphasizing that the product is "positioned as emotional companionship, not a partner." Meanwhile, Unitree publicly questioned that embodied AI is "fundamentally not mature yet" today (for comparison, Unitree's H1 battery is 0.864kWh and Tesla's Optimus Gen2 is about 2.3kWh, with dynamic battery life for both hovering around 2 hours). This controversy follows this week's joint initiative by two industry associations to regulate the development and marketing of "emotional companionship" humanoid robots. Source: National Business Daily source (WeChat, CN)
Hyundai's Atlas Appears at FIFA World Cup to Deliver the Match Ball · humanoid ⚠️ demo/planned figures
Hyundai's Boston Dynamics Atlas appeared at halftime of a World Cup match, delivering the match ball and imitating signature celebration moves of star players, promoting Hyundai's robotics business. Separately, reports say Hyundai plans to produce 30,000 Atlas units annually. The delivery and celebration were staged demonstrations, and the production figure is a planning target — neither is an indicator of scaled deployment. Source: Reuters source
China Institute of Electronics: China's Humanoid Robot Exports Up 210% Year-on-Year · industrial
Xu Xiaolan, chair of the China Institute of Electronics, disclosed that China's humanoid robot exports grew 210% year-on-year, stating that humanoid robots now have preliminary capability to "enter industrial production lines." Separately, industry data shows China's robot exports have this year reached nearly RMB 20 billion, with millions of units shipped to over 150 markets. Source: STCN (Kechuangban Ribao) source (WeChat, CN)
Guangzhou Rolls Out New Policy for Embodied-AI Robot Industry, Targeting 10 Complete-Machine Manufacturers by 2029 · industrial
The Guangzhou Municipal Bureau of Industry and Information Technology released "Measures to Promote High-Quality Development of the Embodied-AI Robot Industry," proposing 8 measures aiming to cluster 10 complete-machine manufacturers and carry out over 15 major technology breakthrough projects by 2029. As a local government policy initiative, this represents concrete industrial organizing action, adding another policy anchor point for the embodied-AI industry in the Greater Bay Area. Source: Guangdong Provincial AI Industry Association source (WeChat, CN)
Supor Signs Strategic Cooperation Agreement with AgiBot · embodied
Supor and AgiBot Robotics have signed a strategic cooperation agreement, launching in-depth collaboration across five directions, with the home-appliance manufacturer further connecting into the embodied-AI industry chain. Source: Cailianshe source
Tsinghua's THU Robotics Team Defends RoboCup 2026 World Championship · humanoid
Tsinghua's THU Robotics team (Booster Robotics-affiliated) successfully defended its world championship in the humanoid soccer competition at RoboCup 2026 in Incheon, extending Chinese universities' lead in embodied motion-control competitions. Source: Sohu source
Unitree's IPO Registration on Shanghai STAR Market Takes Effect, Listing Countdown Begins · humanoid
The Shanghai Stock Exchange shows Unitree's IPO registration on the STAR Market has taken effect, with the market expecting listing to be completed between mid-to-late July and early August — bringing A-share's "first humanoid robot stock" closer to reality (already reported previously). Source: 36Kr source
Hardware · Supply Chain
· Zhongda Leader Intelligent Transmission: As Unitree's largest reducer supplier, accounting for 63% of supply, the company won an exclusive cooperation order worth RMB 3.2 billion total for 2025-2027; its self-developed reducer reaches up to 94.6% transmission efficiency and about 1 arcmin positioning accuracy source (WeChat, CN)
· Yuequan Bionics' Yinshou Y-Hand M1: World's first bionic dexterous hand launch, using a rigid-flexible coupled bionic structure (based on tension-compression body robotics theory), with the full hand weighing a lightweight 299.7g source (WeChat, CN)
· Weihong Co. / Hanmu Electronics: Focused on dexterous-hand finger joint modules, already reaching 70-80 clients at or beyond the pilot-testing stage source (WeChat, CN)
· Color-Changing Tactile Sensor: A new approach lets robots "see" touch through color changes, offering a new direction for low-cost tactile sensing source
Top comments (0)