Today's Highlights

· Amazon now has more than 1 million robots in its fulfillment network, and the number of robotic arms such as Cardinal and Sparrow has doubled this year

· HD Hyundai Robotics uses Palantir's AI agent to modify robot programs, cutting average issue-resolution time from 21.5 days to 7.9 days

· Fysiverse open-sources UniWAM, which learns navigation, docking and hands-on manipulation with a single model

· IT Juzi (Chinese startup-funding data provider) counts 33 embodied AI funding deals in China in October, versus 60 in September

· The CEO of Critical Point says the reliability of dexterous hands with more than 20 active degrees of freedom is not yet good enough for industrial deployment

Fysiverse-UniWAM: One World Action Model for "Where to Go, Where to Stop, and How to Use the Hands"world-model

When a navigation model and a manipulation model are stitched together, a robot often reaches the edge of a table but stops too far away or at the wrong angle, and cannot grasp the object. Fysiverse and Fudan University's CITLab use "manipulation-anchored pose" (MAP, the 2D position and orientation at which the base should stop before the hands start working) to mark docking points explicitly, and automatically generated MAP-Data containing 1.5 million trajectories and about 7,500 hours. The navigation stream and the manipulation stream each keep their own input interface and output head, sharing one video-and-action DiT backbone, without treating large-scale robot pretraining as a prerequisite. The report states that UniWAM-Sim, trained only on MAP-Data with no real-robot data, was deployed directly on Cobot Magic and AgiBot G2 (AgiBot is a Chinese humanoid robot maker), and its navigation arrival accuracy in real environments remained close; success rates on all four mobile manipulation tasks were higher than π0.5, and overall fixed-station manipulation was close to π0.5. The architecture, training and inference code, the navigation weights of UniWAM-Base and Scaled, and MAP-Bench have been open-sourced.

Fysiverse, Fudan University CITLab · Technical report · Source: ITHome (Qudong Zhijia) source

Embodied AI Funding in China in October (IT Juzi Count) | 33 deals | RMB 10.105 billionadjacent

On the same basis, September had 60 deals and RMB 11.028 billion. The number of deals fell by 27 in a month, while the amount fell by only RMB 923 million. Tallies from different sources do not line up: STAR Market Daily (Chinese financial news outlet) counted, on an incomplete basis, at least 185 deals in the third quarter worth about RMB 33.335 billion, with an average of more than RMB 46 billion per quarter in the first half; 21st Century Business Herald (Chinese business newspaper), citing data from Shangqi Chanyetong, put disclosed September funding at RMB 13.161 billion, higher than IT Juzi's September figure. Chen Yonggang recently left his post as vice president of Unitree (Chinese robotics maker) and founded Beijing Guangli Yueqian. In Beijing and Shenzhen, Chen met nearly 20 investors; two are interested in leading a round, three said they would follow if someone else leads, and more said outright that they are "no longer looking at embodied AI." Chen's summary: "It's not cooling, it's already minus 40 degrees." Another institutional source told STAR Market Daily that after attending public physical AI events, some LPs have come asking about the investment risks of this sector. Source: Phoenix New Media Tech (Jiqi Zuiqianxian) source; Huxiu (Cailian Press) source; Sing Tao Global (21st Century Business Herald) source

Dulo (US) | Seed round | USD 130 million | Valuation undisclosedadjacent ⚠️ Unconfirmed report

Founder Sebastian Thrun started Google's self-driving car project. The Next Web, relaying Business Insider, says Dulo is building a foundation model for hardware design, with the goal of "manufacturing at close to the speed of light"; the company has still not launched publicly. The USD 130 million figure comes from the headline of The Information's exclusive report, and Axios the same day called it a large seed round; the company has not confirmed it. Source: Axios source; The Information; The Next Web

Amazon: More Than 1 Million Robots in Its Fulfillment Networkindustrial

More than one million robots are operating in Amazon's fulfillment network, CEO Andy Jassy said on the third-quarter earnings call in response to Morgan Stanley analyst Brian Nowak. CFO Brian Olsavsky added that the number of robotic arms such as Cardinal and Sparrow has doubled this year, and that the company is still developing a robotics foundation model. Jassy said new same-day delivery facilities "rely heavily on robots," with many repetitive jobs going to robots, but that fulfillment centers will "always have lots of colleagues." Jassy gave neither the number of affected jobs nor a timeline, saying only that the current impact is "small compared with the next five years." Amazon gave a different measure in September: more than 1 million robots built cumulatively on its US production lines. Source: Yahoo Finance (Benzinga) source

HD Hyundai Robotics Uses Palantir AI Agent to Modify Programs, Cutting Handling Time by Over 60%industrial

HD Hyundai Robotics says that since August, when it began using Palantir's AI agent for on-site issue handling, average task completion time has fallen from 21.5 days to 7.9 days. The AI agent first reads the customer's existing robot program and proposes changes, and engineers then adjust them based on field experience. In the past this work required a skilled engineer on site, repeatedly modifying work sequences, motion paths and speeds, and it was hard to standardize. CTO Ahn Sung-hwan disclosed the figures in a keynote at Palantir's Sovereignty Bootcamp in New York, and the company plans to train field engineers as "field solution specialists." Palantir Chief Revenue Officer Ryan Taylor said the partnership "starts with work that can deliver measurable impact in the short term." Source: The Korea Times source; Yonhap News Agency source

DEEP Robotics Traffic-Control Robot at the Hangzhou Marathonhumanoid

At the 2026 Hangzhou Marathon on November 1, a humanoid traffic-control robot from DEEP Robotics (Chinese quadruped and humanoid robot maker) took part in event traffic support; the on-site photos published by Global Times were provided by DEEP Robotics. During the National Day holiday, its DR02 had just completed rainy-day duty tests at several intersections in Hangzhou. According to the updated prospectus, the DR-series humanoids have shipped a cumulative 4 units over two years. Source: Shanghai Observer source; Global Times source

Meta Reportedly Built Humanoid Prototype Max in Three Months with Fourierhumanoid

The Information reports that Meta and Shanghai-based Fourier (Chinese humanoid robot maker) built a humanoid prototype, Max, in three months: close to 5 feet tall, weighing about 106 pounds, and running an Nvidia Thor chip. Zuckerberg sees it as the "physical embodiment" of the Muse model and believes it can do everyday household chores. The two companies began work on Max 2 in September; its torso must be redesigned, no launch date is set, and there are no plans to sell the prototype. Reuters said it could not immediately verify the report, and neither Meta nor Fourier has issued a statement. ⚠️ Unconfirmed reportSource: The Information source; ITHome source; Reuters source

Critical Point CEO: Reliability of High-DoF Dexterous Hands Not Yet Ready for Industrial Deploymentembodied

High-end dexterous hands with more than 20 active degrees of freedom "still struggle to meet industrial deployment requirements on reliability at present." Critical Point CEO Qiao Tianjie told China Business Journal and other outlets this is why the company is first bringing low- and mid-DoF products into real-world scenarios. Critical Point was spun off from AgiBot in January this year, and Qiao acknowledged that AgiBot's share of the company's orders is "still relatively high." Qiao expects the research and education market to keep growing for another year or two but with limited room, and says the focus will shift to the deployment market this year and next; dexterous-hand prices may fall by an order of magnitude in the next year or two, provided the market scales up. The newly released OmniGlove Pro data glove has 746 tactile sensing points. Chief Embodied Scientist Huang Bidan explained that "real-robot ready" means a robot can reproduce a demonstrated task under the same conditions, and does not mean it can generalize to unseen environments. Source: China Business Journal (reposted by Sohu) source

Westlake University Establishes World Model Research Centerworld-model

On October 31, the Westlake World Model Research Center was inaugurated in Hangzhou, a university-level interdisciplinary platform at Westlake University (Hangzhou research university) focused on physical intelligence. Director Yu Kaicheng is an assistant professor at Westlake University and also founded Hangzhou Westlake Digital Intelligence Technology. Rather than rendering video or 3D frames one by one, the center breaks scenes down into a finite set of concepts and reasons about object states, spatial relations and causality in an abstract latent space. The team says its pretrained model Awomo has topped multiple international robotics benchmark leaderboards, that related technology has raised over RMB 100 million in cumulative funding, and that downstream developers may need only 20 to 30 hours of post-training to customize a domain model. ⚠️ Vendor claimSource: Zhejiang News source; Pandaily source

Qiyuan Joins Tencent WorkBuddy; Zhishen Signs with Tencent Cloudembodied

Qiyuan Robotics, a unit of Swancor Advanced Materials, became the first embodied AI company to connect to Tencent's WorkBuddy at the Tencent Global Digital Ecosystem Conference. Qiyuan says its PrimeSim world-model simulator topped EWMBench and WorldSimProbe, with a correlation coefficient of 0.99 between simulated policy success rates and those of traditional physics simulators. The same article says the HIL-UMI collection scheme can collect 10,000 data samples versus 100 on real robots under the same budget, and also writes that "overall collection efficiency improved 5.63-fold"; the two multiples do not match. On October 28, quadruped and humanoid maker Zhishen Technology also signed a strategic partnership with Tencent Cloud, under which Tencent Cloud will provide GPU compute, storage and the TI-ONE training platform. ⚠️ Vendor claimSource: Sina (Lei Technology) source; Embodied Home source

Seoul City and Seoul National University to Build Home Robot Test Centeradjacent

The 612-square-meter Physical AX Center will be built at the AI Hub in Yangjae-dong, Seoul, after the Seoul Metropolitan Government signed a memorandum of understanding with Seoul National University on October 30. Participating Seoul National University's H-PAIR institute brings together 41 professors from 16 departments and can receive up to KRW 100 billion in national funding. The test space, replicating a living room and kitchen, will be completed in the first half of 2027, and verification of the performance and safety of caregiving and walking-assist robots will begin in the second half. Seoul National University supports about 5 startups each year. Source: Seoul Economic Daily source

Matrix Robotics to Set Up R&D Center in South Koreahumanoid

Vice President Gao Weize told Seoul Economic Daily that the company aims to set up a joint R&D center in South Korea within a year and to build training-data collection sites with local companies at manufacturing and service locations. The company, founded in 2024 by former Tesla Optimus engineers, says it has accumulated orders for about 1,000 humanoids, and plans to launch its fourth-generation model around May next year. ⚠️ Plan, not yet realizedSource: Seoul Economic Daily source

Robot Mount Tai Climbing Competition Starts November 2humanoid

The course runs from Red Gate to Zhongtian Gate, with 3,100 steps and an elevation gain of 600 meters, with quadruped and humanoid robots competing together. Robots must map the route and recognize terrain in advance. In the qualifying round, "the referees couldn't keep up with the robots," and the format has been changed so that referees are stationed at set points. Source: DoNews source

Retired Philadelphia Professor Arrested for Blocking Delivery Robotadjacent

Carolyn Marvin, a 78-year-old retired communications professor at the University of Pennsylvania, blocked a delivery robot on a Philadelphia street on October 14 and was handcuffed and taken away by police. When she encounters such robots she often blocks them for ten to fifteen minutes, for three reasons: they take work from low-income couriers, they occupy sidewalks, and their cameras invade privacy. "I thought I had more rights than a robot," she said, adding that she will do it again. Uber Eats launched robot delivery in Philadelphia in March, making it its 4th robot-delivery city. Source: CTV News source; The Times of India source