Today's Highlights

· Two humanoid robots collaborated to remove a gallbladder in a live-animal surgery from UC San Diego, published in Nature

· MIT and Motional put an interpretable module on a real car; the same day, another Nature paper clearly explained "phantom braking"

· Travis Kalanick's Atoms has raised $1.7 billion to date, with Uber contributing $100 million

· China's PLA General Hospital used AI ultrasound-robot navigation to complete a transcatheter closure for congenital heart disease

· Supor Aisidi (Chinese cookware/appliance maker) hit two straight limit-up days, though its embodied-robotics business has generated just RMB 137,200 in revenue to date

· LG Electronics is preparing a bipedal humanoid platform, with actuators accounting for roughly 60% of a full unit's bill of materials

· A single video now teaches a new task, as embodied AI starts betting on in-context learning

Two humanoid robots jointly perform a live gallbladder removal · manipulation

Surgical robots have long been specialized, heavy equipment — a typical unit weighs about 1,800 pounds and requires a dedicated operating room, dedicated instruments, and a trained team. Engineers and surgeons at UC San Diego took a different approach: using a general-purpose humanoid robot about 5 feet tall and roughly 60 pounds (which the team calls Surgie), they teleoperated a live gallbladder removal — one procedure performed by a humanoid robot working alongside a human surgeon, and another by two humanoid robots collaborating, in trials on a large non-primate mammal. Senior author Michael Yip said "this research demonstrates a viable future for humanoid robots in surgery," with the value lying in making robotic surgery affordable for rural hospitals and under-resourced regions. The team also flagged unresolved issues: multiple recalibrations were needed during the procedures, operating times were longer than with existing systems, and remote operation remains constrained by communication latency.

Senior author Michael Yip (UC San Diego) · Nature, 2026-09-02 · Reported by VnExpress International source

CW-Net: making a self-driving car's black-box decisions explain themselves — without post-hoc rationalization · autonomy

Most interpretability research stops at simulation and toy settings; this one puts the module into a real vehicle. MIT and Motional's Concept-Wrapper Network is embedded in a middle layer of a machine-learning planner, translating internal reasoning into human-readable concepts such as "approaching a stopped vehicle," and requiring that the final decision be grounded in these concepts — so the explanation has a causal relationship to the vehicle's behavior rather than being a post-hoc rationalization. Training used 130 million labeled driving-scenario samples. A road test made the point clearest: the vehicle repeatedly stopped near cyclists for no apparent reason, and the safety operator assumed the system had recognized the cyclists — but CW-Net's real-time explanations showed the model had not properly configured cyclist detection at all; the stops came from emergency braking triggered at close range. Engineers used this to locate and fix the defect. Lead author Eoin Kenny said that rather than guessing why the car stopped, real-time data provides feedback that lets the system be checked during deployment.

Eoin Kenny et al. (MIT × Motional) · arXiv 2411.18714 source · Nature, 2026-09-02 · Analysis by Netpop source

MetaEarth3D: generating a boundless 3D Earth from a single sentence · world-model

Training data for low-altitude drones and driverless vehicles is bottlenecked by the cost of capturing and labeling real-world 3D scenes. A team from Beihang University pushed generative foundation models from 2D remote-sensing imagery to world-scale 3D scenes: given a passage of text or a satellite image, the system generates an explicit 3D world — from terrain to city blocks — that is continuous, consistent, and boundlessly extendable, with native spatial annotations such as elevation and distance built in, and can be imported directly into UE or Unity. The way around the data bottleneck is to convert 3D scenes into 2D image supervision signals for training. The model has 4.6 billion parameters, was trained on about 10 million globally distributed images, and can generate a district-scale scene of about 17 square kilometers in roughly 2 hours on a single RTX 3090, with measured spatial-understanding gains averaging 22.85%. The team also noted a limitation: current scenes are still fixed to a single point in time and cannot yet show seasonal change or building construction/demolition.

Shi Zhenwei and Zou Zhengxia, co-corresponding authors (Beihang University) · arXiv 2604.22828 source · Accepted by National Science Review · Analysis by DeepTech source

VLAct: the target of VLA pretraining should be the backbone, not the policy · vla

A counterintuitive starting point: after continued pretraining with an OFT head, switching the downstream head to PI actually dropped success rate from 60.5% to 55.1%; switching to a GR00T head dropped it from 51.2% to 28.9%. The paper calls this head-specific representation collapse — more accurate action prediction doesn't mean a more general backbone. VLAct's approach is to attach three continuous action heads for joint supervision, protect the vision encoder and shallow layers, and unify only the action dimensions that are genuinely consistent; all the heads are discarded once pretraining ends. The entire continued-pretraining stage uses only open-source data and 16 GPUs, achieving a 92.5% success rate on RoboTwin 2.0 and 82.6% on LIBERO-Plus, and outperforming every competing model explicitly labeled as a world action model on RoboDojo. The clearest evidence of transfer came from RoboCasa-GR1: the GR-1 robot never appeared even once in the pretraining data, yet using only 20% of the downstream data still beat NVIDIA GR00T N1.6's full-data baseline. The model, checkpoints, training code, and pipeline are all open-sourced.

StarVLA team · arXiv 2608.27550 source · Analysis by JIQIZHIXIN source

Atoms (United States) | $1.7 billion raised to date | a16z-led | Uber makes $100 million strategic investment · autonomy

Nine years after leaving Uber, Travis Kalanick has put autonomous driving back at the center of his company. According to the Financial Times, Uber has invested $100 million in Atoms, and the two sides have held preliminary discussions about Atoms' technology eventually running robotaxis on the Uber platform. Leading Atoms' autonomous driving and robotaxi business is Anthony Levandowski, a Waymo co-founder who later led Uber's self-driving unit, was convicted of stealing Google self-driving trade secrets, and received a pardon from Trump in 2021; Atoms acquired his automation startup Pronto in March, which focuses on automation for heavy industries such as mining. The company was formed when Kalanick renamed his holding entity City Storage Systems in March, then completed a $1.7 billion funding round led by a16z; it now has more than 2,000 employees and plans to keep growing its self-driving team through acquisitions. Speaking at an a16z event, Kalanick said: "A lot of what would have slowly happened at Uber, we're doing now." Source: Financial Times source

VAST | Series B + B+ | roughly RMB 3 billion combined | roughly RMB 5 billion in six months · world-model

Led by Matrix Partners China, with industrial investors including Perfect World, BlueFocus, and Xindongneng joining alongside financial investors CDH VGC, CICC Capital, and CMC Capital; existing shareholders Fortune Venture Capital, Sequoia Capital China (spun off as HongShan), and Oasis Capital added on with oversized follow-on investments. VAST builds native 3D foundation models; its core product Tripo Studio already serves tens of millions of users globally, and Tripo has been integrated into workflows at NetEase, Tencent, ByteDance, and Microsoft. In under six months the company has raised roughly RMB 5 billion in total, a record funding amount in the AI 3D space. The cost of generating 3D assets is an unavoidable piece of world models and simulation training environments. Source: 36Kr Global source

Tiangong Robotics (Beijing Tianxiaxian Zhichuang) | new round | RMB hundreds of millions | investors revealed · industrial

Following its disclosure earlier this month of a funding round worth several hundred million RMB, the investor lineup has now been confirmed: Chuxin Fund led the round, with Matrix Partners China and Hefei state-owned capital participating. The company makes humanoid and intelligent-sorting solutions for logistics applications; the funds will go toward expanding production of sorting robots, R&D and commercial iteration of humanoid sorting robots, and overseas expansion. Global shipments rose from about 2,600 units in 2024 to 16,600 units in 2025, and the company projects growth to 81,700 units in 2026. ⚠️ Company's own figuresSource: 36Kr source

AI ultrasound-robot navigation used to complete transcatheter closure for congenital heart disease · embodied

The patient was a 48-year-old man. The Cardiovascular Medicine Department at the Sixth Medical Center of the PLA General Hospital completed a transcatheter congenital heart disease closure procedure on September 6, during which an AI ultrasound robot system autonomously performed image acquisition, lesion identification, and real-time navigation, guiding precise deployment of the occluder device. The hospital said the system achieved bidirectional intelligent coordination between image recognition and surgical operation, which can reduce complex interventional procedures' dependence on a physician's ultrasound expertise, and stated that an authoritative literature search confirmed this as a global first. Beyond routine cardiovascular diagnosis and treatment, the hospital mentioned potential extensions including battlefield injury assessment, remote robotic surgery, and support for primary-care facilities. ⚠️ Single-party accountSource: Global Times source

Sharp and Toyota Industries cut automated-warehouse outbound-shipment computation to a fifth of prior time · industrial

The two companies announced on September 4 that they had jointly developed a method and computational model using simulated quantum annealing to compute optimal outbound-shipment plans for automated warehouses, and have jointly filed a patent application. Validation was carried out at a large logistics center handling more than 100,000 items per day: conditions affecting outbound routing, such as equipment capacity and shipment deadlines, were first quantified and weighted by importance, then solved using Sharp's simulated quantum annealing technology. Under identical shipping conditions and computing environments, the time needed to derive an optimal outbound plan dropped to about one-fifth of Toyota Industries' original algorithm. The two companies say this can shorten workers' wait times, and next plan to extend the same method to inbound planning and inventory layout. Source: IBTimes JP source

Breeding robot "Jill" unveils embodied-AI version, completing pollination in under 10 seconds · embodied

Holding a pollen tube in its right hand and a pollination brush in its left, the robot dips the brush in pollen and lightly sweeps it across the stigma of a tomato flower, completing a pollination in under 10 seconds, with each dip of pollen sufficient for about 50 pollinations. This version, unveiled in Beijing on September 4, can perform flower recognition, positioning, pollination, and phenotype data collection; the difficulty lies in avoiding branches and leaves while identifying a pinhead-sized stigma among the crop. The other half of what makes this feasible comes from biotechnology: agricultural experts first altered the crop's flower shape so the stigma, normally enclosed, is exposed — eliminating the need to manually pry open petals to find it. "Jill" comes from the team led by Xu Cao at the Institute of Genetics and Developmental Biology, Chinese Academy of Sciences; when the first automated cruising hybrid-pollination breeding robot was unveiled in August 2025, the related findings were published in Cell. Source: CCTV News source

Yuandian Robotics' Bridge "Xiaoqiao" wins IFA Innovation Award, sells 600 units on launch day · humanoid

The Geek Edition launched at RMB 8,888, with the company saying sales topped 600 units on launch day, September 2, and took the top spot on JD.com's humanoid-robot sales chart. The robot won the official award in the "Best in IFA Next & Emerging Technologies" category at IFA 2026, with judge Sascha Pallenberg's citation centered on its open-source design. It ships with bipedal walking and obstacle avoidance built in, offers developers an SDK, motion-control API, and motion-capture/VR access, and comes with OpenBridge, a unified development toolkit and Skill Hub. At this price point, buyers are getting a programmable development platform, not a robot that can perform work. ⚠️ Company's own figuresSource: AI Cloud Information source

Waymo launches invite-only subscription Waymo Premier at $29.99/month · autonomy

Subscribers get 10% back in Waymo Cash per ride, with higher rebates during peak hours, plus priority pickup that skips the queue, early access to new cities, and up to 5 free cancellations per month. Waymo says the idea came from riders themselves. With the nationwide fleet now above 4,000 vehicles (previously reported), a subscription is a step toward locking in ride frequency and retention among high-frequency users. Source: Mashable source

Tesla says Robotaxi will move to 24-hour operation "next month or so" · autonomy

Tesla AI lead Ashok Elluswamy replied this week on X to a user wanting to use Cybercab late at night, saying it would launch once "the next planned v15 technology" is integrated — without specifying which technology, or whether it would first cover Cybercab or the existing Model Y fleet. The service currently operates seven days a week, 6 a.m. to 10 p.m., covering Austin, Dallas, Houston, Miami, Orlando, and Tampa. Low-light nighttime conditions place higher demands on perception and decision-making than daytime operation. ⚠️ Planned, not confirmedSource: IT之家 source

Dubai company and Jiangsu Yunmu Zhizao sign five-year agreement, contracted capacity up to 10,000 humanoid robots per year · humanoid

The agreement was signed in Taicang, with Dubai's No Competition Infinity Gaming Development Services partnering with Jiangsu Yunmu Zhizao; the initial assessment covers scenarios in construction, real estate development, gaming, education, and technical training. On the 10,000-unit figure, the press release itself states it will be adjusted as the project progresses, technology is validated, and commercial applications expand, and that it "does not represent confirmed orders or deployment volumes." ⚠️ Planned figure, not confirmedSource: Fast Company Middle East source

Learn from a single video, embodied AI starts betting its chips on in-context learning · world-model

Skild AI's S1 watches a single human demonstration video and, without updating its weights or any post-training, attempts new tasks it was never trained on — up to ten minutes long and dozens of steps, with tasks like flower arranging, pancake flipping, coffee brewing, and part assembly all outside its pretraining data. The numbers deserve a close look: in a comparison against 100,000 hours of pretraining, video-demonstration prompting scored 66% versus 9% for language prompting; but this score is an average of cumulative per-step success rates across tasks and includes human intervention to rescue failures. Skild's own estimate is that matching single-video performance through additional training would require about 380 teleoperation demonstrations, equivalent to 50 to 100 hours; given 2,000 demonstrations, additional-training performance would rise to 86%, surpassing the single-video approach. Around the same time, Generalist AI's GEN-1.5 also made one-shot learning a core capability, and slightly earlier, Fei-Fei Li, Jim Fan, and Yuke Zhu's RoboTTT was the first to systematically bring in-context scaling to robotic visuomotor policies. A Chinese startup is already following suit: Coco Matrix, founded in April this year by Gao Yuxiang — a graduate of Xi'an Jiaotong University's youth program who dropped out of a PhD at Johns Hopkins and previously led the team at Fourier Intelligence (Chinese humanoid robot maker) that connected the full pipeline from teleoperation to deployment on full-scale humanoids — is taking a different path from the two companies above, moving in-context learning from post-training to pretraining. The marginal returns on stacking more data are declining, and this is a collective industry shift to a different scaling dimension. ⚠️ Company's own figuresSource: QbitAI source, The Rundown AI source

LG Electronics preparing a bipedal humanoid platform, folding robotics into group-wide division of labor · humanoid

"You may have seen our legless mobile CLoiD at CES, but we're also preparing a robot platform with humanoid legs," said Baek Seungtae, head of LG Electronics' Home Appliance & Air Solution business, at a press roundtable during IFA 2026. His stated timeline puts factories first and homes later: humanoid robots will soon be deployed at the company's Tennessee plant in the US and the Changwon smart factory; home environments involve more unpredictable situations and require thorough safety consideration, but "the timeline for robots entering homes could come sooner than people expect." Division of labor has already been set at the group level: LG Innotek handles eye modules and sensors, LG Chem handles materials, LG CNS handles data, and LG Electronics handles manufacturing of components such as joints. Asked about low-price competition from Chinese appliance makers, he assessed that Chinese companies' cost reductions had been running around 10%, but in the first half of this year costs had actually started rising, so the low-price strategy would not persist. ⚠️ Remarks made at an eventSource: Asia Economy source

Supor Aisidi hits two straight limit-up days; embodied-robotics business has generated RMB 137,200 total · humanoid

Supor Aisidi (002403.SZ) issued a trading-anomaly disclosure. Its core business remains cookware, small appliances, and industrial robots; embodied-AI robotics is a newly opened, still nascent business line, having generated sales revenue of RMB 137,200 to date, with technology iteration, project deployment, and commercialization all carrying uncertainty; the company expects no material impact on this year's financial condition or operating results. The cooperation agreement was signed on July 6, 2026: its wholly owned subsidiary, Supor Aisidi Humanoid Robotics Co., Ltd., signed a strategic cooperation agreement with Agibot (Chinese humanoid robotics startup) covering a broad scope — purchase orders and product delivery, contract manufacturing, supply-chain cooperation, technical support, equity investment, and joint-venture establishment — with some elements of the cooperation still in early stages. There is a gap of two orders of magnitude between the agreement's scope and revenue actually realized. Source: Gelonghui source

Unitree robots narrate for Li Auto vehicles, as guided tours remain humanoid robots' first profitable job · humanoid

A humanoid robot sits inside the car to demonstrate cabin space, while a robot dog assists in showcasing smart summon and the trunk — this Li Auto L9 Livis test-drive video opens by specifically noting all footage is unstaged. The collaboration doesn't stop at video: Unitree's "Bengbeng" robot already serves as a guide at the Beijing Automobile Museum, and according to Li Auto's published schedule, it will next perform stunts at Zhuhai Chimelong and give vehicle walkarounds at the Zhengzhou East Station experience center. Figures from Unitree's IPO prospectus explain why such gigs matter: in the first three quarters of 2025, research/education accounted for 73.60% of Unitree's humanoid-robot revenue, commercial/consumer use for 17.39%, and industrial applications for 9.01% — and within industrial applications, corporate guided tours alone made up roughly 50% to 70%. In 2025, Unitree's overseas revenue was about RMB 732 million, or 43.65% of total revenue. Once humanoid robots leave the lab, showrooms, trade shows, and brand events remain their most advanced use case — the tasks aren't complex, but they're generating the first wave of commercial orders. Source: Cross-Border Reference source

About RMB 43.8 billion flowed into embodied AI in H1, over half toward "robot brains" · world-model

According to a compiled report, total embodied-AI funding in the first half of 2026 reached about RMB 43.8 billion, with more than half directed toward "brain" technology, of which VLA models alone captured 42% of the brain-related share. In the same period, passenger-vehicle sales in China fell 24.3% year-on-year to 8.288 million units (China Association of Automobile Manufacturers), while Guo Shougang, head of the Equipment Industry Department at the Ministry of Industry and Information Technology, disclosed at the World Intelligent Connected Vehicles Conference that L2-level driver-assistance penetration reached 70.5%, with navigation-assist penetration rising to 34.2%. As motion control matures and hardware bodies become commoditized, capital is flowing upstream toward wherever pricing power sits. Specific paths are already diverging: Ant Group's LingBot-VLA (embodied-AI unit) 2.0 adapts to 20-plus robot configurations across 17 manufacturers, pursuing a general-purpose base model; ModelBest (Chinese AI foundation-model startup) and Geely's RoboHarness bets on edge-side real-time decision-making, already running routine warehouse operations at a Geely factory; LingYu Intelligence's robots have entered JD.com's 7Fresh supermarkets, sustaining stable operation for more than 8 hours a day. ⚠️ Compiled figuresSource: Sina Technology source

Chinese investors to build humanoid and industrial robotics industrial park in Ussuriysk, Russia · industrial

Kangji International Investment and KRDV Primorsky Krai, a subsidiary of the Far East Development Corporation, signed the agreement during the Eastern Economic Forum, planning to build the "Kangji Zhilian" technology park based on the existing "Kangji" industrial park in Ussuriysk, with investment of 3 billion rubles, including production workshops, research labs, and office space. Beyond finished robots, the Chinese side will also develop robot-control chips and software, and plans to apply for free-port resident status in Vladivostok's free port zone to access tax benefits. Alexey Dunaev, head of KRDV Primorsky Krai, called it a step toward forming a high-tech cluster in the Russian Far East, with the Russian side hoping the project can be adapted to local production processes and Far Eastern climate conditions. Source: www1.ru source