DEV Community

Shawn
Shawn

Posted on

FutureX · Physical AI Daily — Issue 105 (08/31)

Today's Highlights

· AgiBot (Chinese humanoid robotics startup)'s humanoid robot factory built with Minth Group in Serbia begins production, targeting over 5,000 units a year

· Sharpa's dual-arm robot enters a Shanghai DQ store, making a Blizzard in 55 steps

· Pony.ai signs South Korea's FutureLink, plans to deploy 200 L4 vehicles in Seoul, with certification and commercialization timeline still undecided

· China's industrial robot output reached 635,000 units in the first seven months, up 28.5% year-on-year

· Geek+ (Chinese warehouse robotics company)'s new orders rose 35.5%, while its stock price fell nearly 40% below its IPO price

· An OpenAI executive puts affordable L5 at 5 to 10 years away; Musk replied with a clown emoji

Research Progress

Open Source · Tools · Benchmarks

· Bifrost Manifold: This policy evaluation platform for robotics researchers is now open for early-access applications. It can run thousands of rollouts simultaneously on GPUs — a LIBERO sweep that used to take 8 hours now finishes in 30 minutes⚠️ vendor claim. The platform's agent also automatically clusters failure cases and ranks them by impact; models like OpenVLA, GR00T, π0, and Octo can all be evaluated on LIBERO, RoboCasa, RoboMimic, and CALVIN, with replay footage rendered as 3D visualizations via Rerun. (Bifrost AI)

· Microduck: Hugging Face and Pollen Robotics have jointly opened preorders at $399. The bipedal little robot stands just 25 cm tall and weighs under 800 grams, packed with 15 motors, a camera, and a small lidar unit, paired with a 2600mAh battery good for about an hour of runtime; it will ship before Christmas. It doesn't speak, interacting only through sound, and is positioned as an open-source desktop robot with reinforcement-learning-trained posture control. (Engadget, TechCrunch)

· Model Hardware Standard: Anthropic has released a shared specification allowing AI agents to safely operate physical devices, now open for research preview (previously reported). source

Funding & Deals

LatentVerse (潜界科技) | Seed round | hundreds of millions of RMB · world-model

The list of investors in this round has finally been revealed: Hillhouse Ventures, Clearvue Partners, and Innoangel Fund are all on it, alongside two leading embodied-AI companies, AgiBot and Robot Era (Chinese humanoid startup). The round was first disclosed on August 11, and the company itself is less than two months old, with a core team drawn from Tsinghua University's Institute for Interdisciplinary Information Sciences. Founder Hu Yucheng is a student of Robot Era founder Chen Jianyu. The core model they're developing is called UTAM — Unified Tactile-Action Model — which folds intent understanding, world-state prediction, robot action generation, and tactile feedback all into a single embodied-native foundation model trained together. Less than two months elapsed from the start of fundraising to closing, and several VCs have already secured spots in the next round.Source: China Venture source

Psi Robot (灵初智能) | New round | over $100 million · embodied

The investors behind this week's disclosed nine-figure-dollar round have now been named: Tuopu Group, Chery Holding's Ruicheng Fund, Lens Technology, 37 Interactive Entertainment, Wuhu Municipal Investment Holding Group, and Fosun Chuangfu are all in, with existing shareholder Zhuhai Science & Technology Industry Group adding significantly to its stake. The company's proprietary Psi-SynEngine embodied data-collection engine has already been validated at small scale in logistics customers' warehouses. Following this round, core investment will shift more toward world-model research and development.Source: Cailianshe VC Desk source

Sharpa | Cumulative funding | over RMB 4.5 billion | post-money valuation RMB 22 billion · embodied

The post-money valuation figure has now been confirmed at RMB 22 billion. Corporate backers include Alibaba, Meituan, Tencent, JD.com, and Transsion, alongside institutional investors Sequoia China, Qiming Venture Partners, Meituan Longzhu, and Guanghe Ventures. The company was founded in late 2024 by the three founders of Hesai Technology (Chinese lidar maker) and had never previously disclosed its fundraising history.Source: STCN source

Weekly private-market roundup | Aug 22–28 | disclosed funding totaled approx. RMB 16.15 billion · adjacent

According to Cailianshe VC Desk's tally, 136 funding events occurred this week, with advanced manufacturing accounting for the largest disclosed total at approximately RMB 6.625 billion. XPeng Robotics' first-round raise of over $900 million was the single largest deal of the week (previously reported). By round stage, seed/angel rounds led in count with 49 deals, about 36% of the total; Series A rounds, though only 36 in number, accounted for the highest disclosed total, roughly RMB 8.215 billion. By region, Shanghai led with 31 companies funded, followed closely by Shenzhen with 26.Source: Cailianshe VC Desk source

Commercialization & Deployment

AgiBot and Minth's humanoid robot factory begins production in Šabac, Serbia

On August 29, Serbian President Aleksandar Vučić personally toured the production floor in Šabac and, in a lighthearted moment, played goalkeeper against a shooting robot. The factory was built through a partnership between AgiBot Innovation, Minth Group, and Serbian partners, and is the first robot factory built in Europe through Chinese corporate investment. In his remarks, Vučić thanked China and noted that Minth has already created nearly 5,000 jobs in Serbia, saying the country "is becoming a pioneer in Europe's scaled mass production of humanoid robots." He continued: "We hope that in the future, this factory will be able to assemble and produce more than 5,000 humanoid robots a year." These robots are intended for use in healthcare, agriculture, industry, and defense. The first batch, rolling off the line in 2026, will initially be deployed at the 2027 Belgrade Specialized Expo, with full-scale mass production not expected until 2027. The figure of 5,000 units a year remains an expectation stated by the president, distinct from the production line's current actual capacity.Source: Guancha source

Sharpa's dual-arm robot goes to work at a Shanghai DQ store, operating 12 hours a day

From cup retrieval to mixing to inverted delivery, the entire 55-step process runs without any human intervention. Starting August 29, a white dual-arm robot named "Xiao Sha" has been working at a DQ store on Wujiang Road in Shanghai, taking about 6 minutes to make a Blizzard, compared to two to three minutes for an experienced human staffer. The store operates from 10:00 to 22:00 year-round, with the robot counter and the human counter running as two parallel lines; online reservation slots are already booked out several weeks in advance. Sharpa's first principle for the project was never to modify the environment: the equipment, ingredients, processes, and tools all remain exactly as they were, with the robot handling the same paper cups and mixing spindles used by human staff. The company developed a proprietary hierarchical end-to-end model called CraftNet that incorporates tactile input, paired with the Sharpa Wave dexterous hand; 98% of the 55 steps directly rely on tactile sensing — for instance, sensing friction when pulling a cup, or continuously adjusting grip force in response to resistance during high-speed mixing. Co-founder Li Yifan acknowledged that the first-generation robot has not yet turned a profit, and that this early wave of deployed scenarios industry-wide has very little chance of positive ROI in the short term. He cited three long-term variables: whether the robot's operation can approach human-level performance, whether the hardware can last more than 10 years, and whether mass-production costs can drop from the hundreds of thousands of RMB range to around RMB 100,000.Source: STCN source

Pony.ai signs FutureLink, to deploy 200 L4 vehicles in Seoul by 2028

An initial batch of 10 vehicles will go through South Korea's certification process first, with the remaining 190 to follow once certification is complete. On August 28, Pony.ai signed a partnership in Seoul with South Korean mobility company FutureLink. The vehicle is based on BAIC's Arcfox Alpha, with the seventh-generation autonomous driving system integrated directly on the vehicle production line rather than retrofitted afterward, and redundancy built into critical functions like braking, steering, and power. Service will initially launch in Seoul's Gangnam autonomous driving pilot zone, before expanding to the rest of Seoul, the greater metropolitan area, and other major cities; specific certification and commercialization timelines have not yet been set, and the two companies are also considering forming a joint venture. FutureLink said it has already conducted day-and-night testing in Gangnam using a Kona Electric fitted with Pony.ai's kit, covering roughly 80,000 km without incident. Founder Peng Jun told The JoongAng Daily that South Korea was chosen because of its large automotive industry scale, government willingness to improve regulation, and relatively high taxi fares: "Seoul's high urban density and complex road conditions make it one of the more technically demanding regions for autonomous driving — we see South Korea as a proving ground for the technology." He also noted that operations in Guangzhou and Shenzhen are now profitable. The production cost of the seventh-generation kit is 70% lower than the previous generation, and the company plans to cut costs by a further 20% in 2026, aiming to bring the combined cost of kit and vehicle below RMB 230,000 by mid-2027. In Shenzhen, the fleet averaged 23 paid rides per vehicle per day this past February.Source: The Korea Times source, The JoongAng Daily source

South Korea's Jeonbuk National University builds a physical AI pilot lab to add "autonomous judgment" to automated production lines

The 846-square-meter lab houses 9 robot cells and was built last year through a KRW 6.77 billion proof-of-concept project. A robotic arm places steering wheels onto inter-process transfer equipment, another arm continues processing and assembly, and AMRs shuttle materials between cells. When a reporter accidentally stepped into a robot cell's safety zone, an alarm sounded immediately and the equipment stopped, with the digital twin display on the wall syncing the updated status about 0.5 seconds later. The lab has separate process configurations for three product types — steering wheels, mirrors, and others — with each cell assigned its own agent responsible for that process, plus an upper-level integration agent overseeing the whole system. When sensor readings go out of range, the cell-level agent flags the anomaly while the upper-level agent assesses its scope and offers recommendations to workers. Currently the system only handles anomaly detection and response suggestions — actually adjusting equipment or resuming a process still requires the main control system. In other words, if a piece of equipment fails and the issue touches process chains involving other facilities or logistics robots, human or upper-level control system intervention is still required. Kim Sun-tae, head of Jeonbuk National University's Physical AI Convergence Technology Program, said: "Putting humanoid robots into the mix won't immediately raise factory productivity." In his view, what really matters is a factory operating system that lets existing equipment coordinate autonomously — not the robots' movements themselves.Source: Digital Today source

Singapore Post trials autonomous logistics vehicle in Tampines, running only 4 to 7 a.m.

The vehicle being trialed is the Zelos Z10, shuttling between SingPost's regional e-commerce logistics hub and the T-Space building on Tampines Concourse Link, equipped with 360-degree sensors, though a safety operator still rides along on each trip. The news was announced late on August 29 via Facebook by Ma Yam Keng, Mayor of North East District and MP for Tampines GRC. The trial hours are timed to avoid daytime traffic peaks entirely.Source: 8world source

A 2.3-mile Tesla Robotaxi trip is quoted at 66 minutes

A passenger identifying as a "Tesla Ambassador," Reggie Overton, hailed a ride in Austin for a 2.3-mile trip, only for the system to estimate 1 hour and 6 minutes. He initially thought it was a software bug and called customer service; after several rounds of back-and-forth he was told the route couldn't be recalculated, and he ultimately had to change his destination just to end the ride, then hailed a separate Uber. Tesla later refunded him. He wrote on X: "For a 2.3-mile trip, this experience was extremely frustrating, especially since customer service couldn't correct an obvious routing problem." Tesla's Robotaxi service currently remains limited to Austin, Dallas, and Houston in Texas, and Miami, Orlando, and Tampa in Florida. The invite-only launch event for the Cybercab is set for September 3. Musk said months ago that he expects robotaxi service to be available to about half the U.S. population by year-end.Source: Chron source

Industry Developments

China's industrial robot output hits 635,000 units in first seven months, up 28.5% year-on-year

According to data from the China Machinery Industry Federation, cumulative industrial robot output for the first seven months reached 635,000 units, up 28.5% year-on-year, while cumulative power generation equipment output over the same period reached 208.97 million kW, up 2.8%. Value-added output grew year-on-year across all five major machinery-industry categories: instrumentation manufacturing grew 9.4%, specialized equipment manufacturing 8.8%, general equipment manufacturing 7.9%, automotive manufacturing 7.2%, and electrical machinery and equipment manufacturing 6.5%.Source: People's Daily source

Geek+'s new orders rise 35.5% while stock falls below IPO price

New orders in the first half reached RMB 2.385 billion, up 35.5% year-on-year, nearly 1.9 times revenue for the same period; revenue came in at RMB 1.284 billion, up 25.3%, with gross margin rising from 35.1% to 35.8%. Revenue from outside mainland China accounted for over 75% of the total, with overseas sales carrying a gross margin of 46.2%. Subscription-service orders reached RMB 156 million, up over 75% year-on-year, with the Americas region up as much as 455%; orders in manufacturing scenarios grew over 600%, while the pallet-to-person product line grew over 200%. Adjusted net loss was RMB 60.6 million, narrowing 32.1% year-on-year; the company also noted that excluding RMB 44.5 million in embodied-AI R&D spending, the adjusted loss would have narrowed by 81.9%. However, the reported net loss widened from RMB 50 million in the same period last year to RMB 180 million, driven mainly by a RMB 100 million foreign-exchange loss. AIX Finance noted that net cash outflow from operating activities reached RMB 290 million in the first half, versus just RMB 110 million a year earlier, while receivables and notes receivable rose from RMB 960 million to RMB 1.11 billion. On embodied AI, the company disclosed only that "orders have achieved a breakthrough," without giving specific figures, shipment volumes, or revenue contribution; co-founder Chen Xi said in July that the Gino 1 is scheduled to enter small-batch production and delivery in the third quarter of this year. The stock closed at HK$10.2 on August 28, below its HK$16.8 IPO price and down nearly 70% from its January high.Source: Jingbao Net source, Chaoqi Net source

OpenAI product VP draws a falsifiable line for "affordable L5"

Peter Welinder wrote a lengthy post on X arguing that affordable L5 autonomous driving is still 5 to 10 years away, on the grounds that it requires AGI, which doesn't yet exist: first there needs to be multimodal, multiplexed AGI, then real-time inference capability, and finally all of that has to fit onto a power-efficient, inexpensive chip that can go into a car ordinary people can afford. Rather than stopping at the assertion, he laid out concrete thresholds: the vehicle should cost under $30,000, or a robotaxi's cost per passenger-mile should be under $0.65; there should be no geofencing, no remote operators, and no reliance on GPS or centimeter-level HD maps; the system needs to handle weather, faded lane markings, construction, and police hand signals, and cope with real-time negotiation at four-way stops in cities where lane markings haven't been repainted in a decade; and it needs to be able to recover from problems on its own, with no human takeover required. Musk's only reply was a clown emoji — no quote, no argument, no data.Source: OfficeChai source

XPeng's third Turing chip lights up in production vehicles, completing its three-chip architecture

Following the second-generation VLA's integration of 30-second memory and 6-second prediction into the model (previously reported), XPeng announced on August 27 that the third of its proprietary Turing-series chips has now lit up in production vehicles, with the three chips handling perception, decision-making, and cross-domain fusion respectively. Since mass production of the Turing chip began in Q3 2025, cumulative shipments have exceeded 200,000 units; the company says it will switch its entire lineup to the proprietary chip starting in Q2 2026, targeting annual shipments approaching one million units. The "super-agent" running on this hardware is a new version formed by merging the second-generation VLA and VLM, with an on-device model parameter count 3.5 times larger than the previous generation. End-to-end response speed is up 300%, enabled by streaming inference, in which the model receives input, reasons, and outputs a driving trajectory simultaneously rather than waiting for full processing to complete.⚠️ vendor claimSource: GlobalChinaEV source

Sanctuary AI starts selling its "robot brain" separately, with the humanoid body taking a back seat

This summer, on the production line of a global Tier 1 automotive parts supplier, the company plugged flexible harness connectors into moving targets with a reported success rate above 99.5% and a cycle time of 2.54 seconds. The footage shows no humanoid robot — just two suspended hands grasping wire, rotating, and inserting connectors. New CEO Daniel Friedmann explained how the figure was derived: testing ran for 40 minutes across 313 insertion attempts. The company's homepage now leads with "physical AI for industrial automation," and while the Phoenix Gen 8 humanoid robot still appears on the site, it now ranks below robotic arms, dexterous hands, and grippers. Rather than waiting for humanoid hardware to reach large-scale commercial deployment, the company is putting its models into existing industrial workflows and robots first — a hardware-agnostic approach that lines up well with Meta's real-world need for robots to plug in network cables and restart servers in data centers. The 40-minute, 313-attempt figure reflects only the scale of a capability demonstration, not proof of annualized production-line reliability.Source: Forbes source

Thirty LG executives fly to Santa Clara as physical AI partnership moves into technical talks

About 30 executives and business leaders flew to Nvidia's Santa Clara headquarters for a two-day session, holding multiple rounds of technical meetings on humanoid and logistics robot development while assessing commercialization feasibility. Under the plan announced on June 8, LG Electronics will use Nvidia's Isaac, GR00T, and Cosmos platforms to build out the full pipeline from data collection and simulation training to model deployment; on the manufacturing side, LG's production-line data will be connected with Nvidia's computing power and digital twin capabilities. The delegation included LG CNS president Hyun Shin-kyun, LG Electronics CTO Kim Byoung-hoon, and LG Innotek CTO John Min. News of the trip sent LG Electronics and LG CNS shares up as much as 14% intraday, with LG Corp also rising about 7% at one point. Yonhap reported that specifics of the first projects are expected to be announced only in a few months.Source: TradingKey source

Asus creates a Physical AI business group, renaming its AIoT unit on September 1

Asus's AIoT business group is being restructured and renamed the "Physical AI Solutions Business Group" (PAS BG) starting September 1, co-led by Jonney Shih (Zhang Quande) and Samson Hu (Xu Youjia), and organized into three divisions: AI Solutions, Autonomous Platforms, and Humanoid Robotics. Co-CEOs S.Y. Hsu and Vivian Hu wrote in an internal memo that as AI moves toward real-world systems, the company aims to stay ahead in the next wave of innovation. Asus estimates the total physical AI market will reach $1.1 trillion by 2035.⚠️ vendor claimSource: Taiwan Economic Daily source

First-half humanoid robot sales ranking: UBTECH, AgiBot, Unitree lead the pack

This ranking comes from the media account "Humanoid Robot Scenario Application Alliance," covering full humanoid units — bipedal, wheeled, and other mobile configurations, from full-size to compact — with figures estimated from a combination of public company data, historical production and sales trends, and market information. UBTECH led with 16,123 units sold in the first half (previously reported), followed by AgiBot in second place, which rolled its 15,000th unit off the line on June 28; Unitree came in third, saying in early June that cumulative production of its bipedal humanoid had reached about 11,000 units as of May. These top three companies are moving from the thousand-unit threshold toward tens of thousands, while most other companies remain at the small-batch validation and project-delivery stage.⚠️ compiled estimateSource: HK01 source

Hardware · Supply Chain

· Leaderdrive (Chinese harmonic reducer maker): plans to issue H-shares for a Hong Kong listing (previously reported); first-half 2026 revenue reached RMB 349 million, up 38.64% year-on-year, with net profit attributable to shareholders of RMB 70.1092 million, up 31.25%. According to QY Research, the company held about a 12% share of the global harmonic reducer market in 2025, ranking second worldwide; its share of the China market was 27.5%, ranking first. source

· Gaoteng Robotics: its Mecanum-wheel humanoid embodied chassis has already begun partnerships with several embodied-AI companies, with the company disclosing at the Shenzhen AGIC expo that shipments have reached 200 units⚠️ self-reported at trade show. source

Top comments (0)