Today's Highlights
· Galbot (Chinese embodied-AI startup) issued a statement calling online claims about the company false, saying it has filed a police report, escalating the data-collection-loop dispute after Mech-Mind CEO Shao Tianlan named it directly
· Suzhou-based Lexoo Technology alleges OpenAI copied the conceptual framework and website design of its Aether large model, and has launched litigation
· Hyundai's IONIQ 5 deliveries to Waymo are set to begin in Q4 this year, with vehicles rolling off the Georgia plant line before being fitted with the sixth-generation Waymo Driver
· In Hefei, 19 humanoid robots are patrolling on a dedicated 5G-A network, alongside China's first communications plan launched specifically for embodied robots
· ShengShu Technology's Motus2 integrates an evaluator into its world model, achieving an 84% average success rate across five real-robot dexterous manipulation tasks
Paper Progress
Motus2: giving world models an interface for "was this result any good?" · world-model
World models previously answered only two questions — what action to propose and what would happen after — missing the step of scoring the outcome. Motus2 folds a value model into the same set of parameters, letting the robot generate N candidate actions internally, simulate their consequences, and execute the highest-scoring one. Across five real-robot tasks — placing a small ball, multi-finger manipulation, fitting a rubber piece, screwing in a light bulb, and placing a phone — the average success rate was 84%. Looking just at placing a phone and multi-finger manipulation, the base policy scored 65%, model-based reinforcement learning raised it to 72.5%, and adding inference-time planning pushed it to 75%. ShengShu Technology co-founder and CEO Luo Yihang showed a real-robot demo at the Bund Summit, including two-handed paper tearing, page turning, and dough rolling.
ShengShu Technology · arXiv 2608.30237 source · Analysis: AI Era source
JoyAI-EchoWM: an explorable world you can hear, released open source · world-model
Interactive world generation used to be judged on visuals; now it's judged on whether the door you left through is still where you left it when you turn back. JD Explore Academy has put native joint audio-video generation and first-/third-person camera control into a single framework, mapping WASD discrete key presses into continuous 6-DoF trajectories via a unified "camera intent," with sound tracking the visuals in sync. Across 158 WBench Navigation cases, it averaged a score of 81.7, with consistency at 89.8 and interactivity at 87.2; the four-step causal streaming variant, EchoWM-Flash, averaged 81.0 with interactivity at 87.9. The team acknowledges the model lacks explicit persistent 3D memory, so geometric relationships and object identity can drift after multiple rounds of generation. Code is up on GitHub.
JD Explore Academy · arXiv 2608.23189 source · Analysis: JIQIZHIXIN source
EMERGE-Policy: no longer letting one policy decide everything · manipulation
In long-horizon tasks, a single end-to-end model handling low-level perception, high-level planning, and fine-grained action output all at once accumulates error until it collapses. Professor Li Xiu's team at Tsinghua Shenzhen International Graduate School packaged VLA, motion primitives, world models, and verifiers all into standardized tool interfaces, handed off to a main agent and role-based sub-agents for asynchronous scheduling, and used a three-layer file-based memory system — PLAN.md / HISTORY.md / MEMORY.md — to sustain extremely long horizons. On the standard LIBERO benchmark, average success rate reached 99.2%, or 98.8% when using pi_0.5 as the execution skill; on the LIBERO-Plus perturbation test with lighting changes and sudden camera pose shifts, it scored 93.9%, 11.7 percentage points higher than a single Cosmos Policy backbone.
Fang Zhirui et al. (Tsinghua Shenzhen International Graduate School) · arXiv 2608.29896 source · Analysis: Sina Technology source
A humanoid robot swings across a horizontal bar using raw lidar point clouds · locomotion
A thin bar suspended in mid-air, with sparse geometry, is a scenario that stalls perception and whole-body control simultaneously. ETH Zürich's Robotic Systems Lab had an EngineAI PM-01 jump onto a bar, brachiate hand-over-hand, and land in one continuous motion, succeeding in 14 of 15 hardware trials across three bar configurations, with bar heights of 1.69 to 1.75 meters, spacing of 0.26 to 0.33 meters, and peak swing speed of 0.5 m/s — in one trial the rig itself was still swaying. Raw point clouds from a head-mounted RoboSense E1R solid-state lidar are fed directly to the controller, with a GRU memory tracking the bar's position after it moves out of view. Training happens in two stages: first training separate expert policies for jumping on, swinging, and landing, then distilling them into a single student policy.
ETH Zürich Robotic Systems Lab · Learning Agile Perceptive Traversal of Sparse 3D Structures for Humanoids · Reported by: TechEBlog source
Other papers today: finishing a task isn't enough — a benchmark separately measures generative agents' resilience and collaborative judgment under accumulating technical, human-factor, and operational disruptions (arXiv 2609.10724 source); on which updates should be allowed through for continually learning embodied agents, the authors show interval-based confidence gates fail to certify that old-task behavior is unchanged even within a fairly large interaction budget (arXiv 2609.10873 source); a developmental-psychology framework for how autonomous agents acquire and align with social norms (arXiv 2609.11660 source).
Funding & Deals
Mecka AI｜New Round｜Amount Undisclosed｜Valuation ~$500M · adjacent ⚠️ unconfirmed reporting
Led by Sequoia Capital, with the valuation disclosed to TechCrunch by two people familiar with the matter; round size is unknown and terms have not been finalized. This comes just three months after its previous round, led by Framework Ventures at $60 million, with participation from Menlo Ventures, SV Angel, and Kindred Ventures. Founded in 2024 by four co-founders none of whom have a robotics background, the company's approach is to pay people to wear motion sensors and film themselves on smartphones doing everyday tasks like making coffee or fixing a car, then sell the first-person human data to robotics companies and AI labs. Co-founder Josh Gao told Fortune in June that the company expects to reach $100 million in annualized revenue by the end of 2026. A rival in the same space, XDOF, was reported last week to be in talks for a new round at a roughly $1.2 billion valuation.Source: TechCrunch source
Skyverge Intelligence｜Angel+ Round｜Hundreds of Millions of RMB · world-model
Participating investors include Linear Capital, Junshan Capital, Hongyi Capital, and Yinshan Capital. The company was founded only this past February in Hangzhou; founder Chen Wei is Li Auto's former Chief AI Scientist and head of its foundation model division, and co-founder Zhang Xiao is Li Auto's former President of Product Lines. Its technical approach, called Duplex Reasoning, centers on making dialogue and action a bidirectional channel — instructions can be interrupted and corrected mid-course, and goals can be adjusted during execution, rather than following a simplex "receive instruction, execute, return result" flow. The commercial rollout plan is to start with hotels and elder-care facilities before entering homes, prioritizing high-frequency, long-horizon tasks like laundry, tidying, and cleaning. Chen Wei described the approach to this round as "in a sense, building models is about digging in for the long haul and grinding it out."Source: PEdaily source
Zhuhai Amicro Semiconductor｜Hong Kong Listing Filing Approved｜Plans to Issue Up to 90.9114 Million Shares · hardware
The China Securities Regulatory Commission's International Cooperation Department has issued a filing notice, also approving the conversion of a combined 360 million unlisted domestic shares held by 43 shareholders into overseas-listed shares for trading. Founded in Hengqin in 2014, the company changed its name from "Amicro Semiconductor" in October 2025, and focuses on robot main-control chips, drive-control algorithms, and high-precision sensors; historical shareholders have included Xiaomi's investment arm and Gree's investment arm. The company has nearly 400 employees, with R&D accounting for 80%. Its AM970 robot main-control SoC features a 5TOPS NPU and dedicated SLAM IP, supporting six parallel HD video streams. The company states its chip solutions have been deployed in over ten million commercial robots and devices worldwide.⚠️ company's own claimsSource: Robot Insight source
HMS Networks Acquires ROBOT S.A.｜Friendly Offer｜€17.2 Million Consideration · industrial
Swedish industrial communications maker HMS Networks has launched a friendly public offer for all shares of Spain's ROBOT S.A.; its subsidiary HMS Industrial Networks SLU has signed a share purchase agreement.Source: GlobeNewswire source
Commercialization & Deployment
19 humanoid robots run on a dedicated 5G-A network in Hefei; China's first communications plan for robots launched · adjacent
On September 10, China Telecom's Anhui branch, together with LimX Dynamics (Chinese robotics company) and Huawei, held a launch event in Hefei's Yaohai District China Netcity, introducing the "Embodied Wing Connect" plan for embodied robots, offering 500Mbps peak uplink, 20ms latency, and priority service scheduling. On site, 19 humanoid robots fitted with custom 5G-A modules patrolled in parallel across eight main routes in the demonstration zone, each streaming multiple HD camera feeds back in real time simultaneously, with end-to-end latency staying stably below 80 milliseconds even during multi-device operation. Supporting this is a 5G-A high-uplink network recently deployed at scale in Hefei, using an 8T8R configuration, with uplink speeds in core business zones stably above 20Mbps and peaks exceeding 500Mbps. Data collected during patrols flows back to LimX Dynamics' data platform, where it is cleaned and annotated for model iteration.Source: Sohu source
Hyundai to begin delivering IONIQ 5 to Waymo in Q4 · autonomy ⚠️ stated as plan
At its 2026 CEO Investor Day in Seoul, Hyundai Motor said deliveries of the IONIQ 5 to Waymo are set to begin in Q4 this year, with vehicles assembled at the Metaplant America plant in Georgia before Waymo integrates the sixth-generation Waymo Driver. These vehicles aren't simply cars with sensors bolted on afterward — the autonomous version comes with built-in redundant hardware and electric doors. The partnership was first disclosed in October 2024, when both sides said they would build a significant number of vehicles over multiple years, but the Investor Day materials did not give a binding total fleet count. Waymo has already been operating the IONIQ 5 autonomously in San Francisco with a safety attendant aboard; switching the underlying vehicle platform means sensor placement, braking characteristics, steering response, and emergency procedures all need to be retuned as an integrated whole.Source: FOX8 WGHP source
Zoox takes on Waymo in San Francisco with free rides · autonomy
Fleet size is the backdrop to this story: Zoox has about 100 vehicles in San Francisco, versus nearly 4,000 for Waymo. Zoox's vehicles are built from the ground up without a steering wheel, pedals, or driver's seat, with passengers seated facing each other inside a teal-colored cabin. Since last November it has offered free rides in San Francisco, expanding in June to eight restaurants plus landmarks including the Ferry Building, Painted Ladies, and Castro, and opening a Rider Lounge as a pickup point downtown in May. Charging fares still requires clearing both the California DMV and CPUC; the company has submitted a deployment permit application to the DMV, but Zoox is not currently on the CPUC's permit list. VP of Communications and Marketing Carly Wyatt told the New York Times, "We're moving as quickly as we can, but this is a process that takes months," adding, "We recognize we're taking baby steps." Las Vegas is its first paid market, with fares set slightly above Uber and Lyft.Source: TechSpot source
A Cybercab spotted on New York streets has a steering wheel · autonomy
A gold Cybercab photographed by Gothamist had a person in the driver's seat with a steering wheel within reach — inconsistent with the steering-wheel-free, pedal-free version Musk unveiled in 2024. Officials confirmed these drivers are testing a "supervised" autonomous driving system and must be able to take over at any time. Tesla has not yet initiated an autonomous-vehicle testing permit application with New York State's DMV; the activity looks more like data collection than an imminent service launch. Tesla's Robotaxi fleet in Austin currently stands at 45 vehicles; over the same period, Waymo said it completed more than 500,000 paid trips per week in Q1 this year. Tesla declined to disclose how many vehicles are in New York or how long testing will run.Source: Futurism source
UBTECH signs exclusive agent for Egypt · humanoid
UBTECH announced on September 11 that it signed a strategic cooperation agreement in Shenzhen with Misr Elsalam for Development & Advanced Technology S.A.E., which becomes its exclusive regional agent for Egypt, targeting education, business services, and public affairs applications. The agreement did not disclose financial terms or initial order volume.Source: Beijing Business Today source
XPeng announces mass production of IRON humanoid at IFA Berlin · humanoid ⚠️ company's own claims
Following the first IRON unit rolling autonomously off the production line on the 8th of this month, XPeng characterized this at IFA 2026 as the start of mass production. The production version stands about 1.7 meters tall and weighs around 65 kilograms, with 76 degrees of freedom overall and 21 in each hand. The company said the robots are being produced at an existing factory but did not disclose planned production volume. XPeng describes its target use case as household chores, specifically naming floor sweeping and folding laundry.Source: Yahoo Tech source
Industry Developments
Galbot files police report as the data-collection-loop business model is put on public display · embodied
On September 12, Galbot (Chinese embodied-AI startup) issued a strongly worded statement saying that content circulating online about the company and its management team is false, that it has filed a report with police, and that it will pursue legal action against those spreading the claims. The trigger was a series of posts on September 10 by Mech-Mind founder and CEO Shao Tianlan on his WeChat Moments, attacking what he called "consortium-style entrepreneurship" — using data-collection centers, leasing companies, and other related-party transactions to manufacture false, unsustainable revenue — and naming Galbot directly in the comments. Galbot responded that evening with a lengthy post titled "Embodied Intelligence Is a Marathon, We're Only Racing Against Time," which did not directly engage with the accusation, saying only that it would "not participate in pointless verbal disputes" and would "not be distracted by short-term noise."
The technical backdrop to the dispute is a severe shortage of real-robot teleoperation data. According to an interview by outlet DingjiaoOne, Lingyu Zhineng CEO Jin Ge laid out the economics as follows: a robot unit sells for roughly 150,000 RMB, a data-collection center buys back data at 120 to 150 RMB per hour, a center running 100 units produces 400 hours of real-robot data per day, and a robot pays for itself in about a year. Another arrangement Shao referenced is embodied-AI companies setting up joint-venture data-collection firms with local governments, with the joint venture funding purchases of robot units, and the embodied-AI company then buying back the data those units produce. One investor summarized this as local governments issuing a five-year "bond," with the company repaying it in installments via data purchases, inflating both revenue and valuation simultaneously. A Hong Kong-listed brokerage analyst said related-party transactions need to be judged on commercial substance, and that revenue from money simply circulating in a loop should be discounted.
According to IT Juzi, 288 funding deals took place in China's embodied-AI and robotics sector in the first half of 2026, involving 226 companies, with disclosed amounts exceeding 46 billion RMB — already surpassing all of last year. China Business Journal reported that as of mid-2026 there were at least 15 large-scale embodied-AI data collection sites nationwide. The secondary market is moving in the opposite direction: Unitree Robotics closed at 477 RMB per share on September 11, with a market cap of 192.978 billion RMB, having briefly touched 444.9 billion RMB on its listing day, August 19; Mech-Mind fell below its issue price on its very first day of Hong Kong trading on September 1, and closed at HK$95 on September 11, below its HK$101.7 issue price.Source: Eastmoney source, DingjiaoOne source
Lexoo Technology alleges OpenAI copied its model narrative, has launched litigation · world-model ⚠️ one-sided account
Suzhou-based Lexoo Technology CEO Guo Renjie published an open letter in Chinese addressed to OpenAI, point-by-point comparing similarities between his company's Aether large model and OpenAI's recent announcements. According to CNBC's translation, he said: "People always say big tech companies have intelligence networks monitoring the entire internet — this time I believe it. This is a direct, unmodified distillation of us." Among the parallels he listed: recursive self-improvement, using AI to optimize compute resources, splitting alignment into goal alignment and value alignment, and a brain-cerebellum-cortex-spinal-cord-inspired architecture modeled on neuroscience; visually, he said OpenAI's GPT-6 Astra website, which went live on September 3 with a dark theme and particle effects, closely resembles the design of the Aether website from two months earlier. Lexoo Technology says its own website went live this past July, and that its founder had presented the "visitor from beyond the stars" framework in person in Silicon Valley weeks earlier. Guo Renjie said the company has begun filing litigation. Recursive self-improvement and using AI to optimize compute are themselves established topics in public research, so it's equally plausible that different teams arrived independently at similar directions and implemented them in entirely different ways.Source: Tekedia source, AI Era source
FCC adds advanced robots manufactured overseas to its Covered List · industrial
The International Federation of Robotics (IFR) has issued a member notice regarding this measure, announced on July 28. It applies to mobile ground robots capable of operating away from an operator, equipped with sensors and network connectivity, covering service robots, AMRs, humanoids, quadrupeds, and some consumer robots; articulated, Cartesian, SCARA, and parallel industrial robots are explicitly excluded. The determination is based on where the product is manufactured, not the manufacturer's nationality — companies headquartered in the United States, Europe, or Asia could all be covered if their products are made outside the United States. Newly restricted products cannot obtain the FCC equipment authorization needed for import, marketing, and sale, but the measure is not retroactive — already-authorized products may continue to be imported and sold, and software and firmware updates remain permitted. The FCC also maintains a conditional-approval pathway, requiring applicants to submit materials on equity structure, investors, supply chain, manufacturing arrangements, software/hardware architecture, and cybersecurity controls.Source: IFR source
Alipay connects AI payments to a robot dog · adjacent
Alipay announced a partnership with Amap Dongliang (Alibaba mapping unit's robotics arm) to launch "AI Pay · Embodied Intelligence," under which Amap Dongliang's robot dog "Tutu" can run errands on command and complete payments within limits authorized by its owner. The underlying infrastructure is Ant Group's APASS commercial trust framework, which applies a KYA (Know Your Agent) approach to verify an agent's identity, intent, authorization, and behavior end to end. Alipay also said it will launch an AI Wallet agent to manage every AI-initiated payment.Source: IT Home source
Hall 3 at the China International Fair for Trade in Services lays out the full embodied-AI supply chain · humanoid
Hall 3, themed around telecommunications, computing, and information services, brought together more than 130 companies and institutions. Tiangong (Chinese humanoid robot, repeat record-setter at the World Humanoid Robot Games), fresh off repeatedly breaking human records at the World Humanoid Robot Games, gave a static display, while another robot struck bianzhong chimes on site, with force control governing the weight and rhythm of each strike. Making its first appearance was a Chinese-developed robot electronic-architecture platform, with a single architecture covering everyday, industrial, and medical scenarios. On the data side, an embodied-AI data services company showed off hand- and arm-worn capture devices that turn grasping, holding, and handing motions into robot training material; its lead, Dong Xiaochao, said the company has amassed a database of over 1.5 million hours of human behavior data and is working to productize the data so the same dataset can be reused across multiple teams.Source: CCTV.com source
Hardware & Supply Chain
· Jiangsu Leili｜Three technical approaches to dexterous hands: a linkage-driven design with 18 degrees of freedom and grip force over 50N, a cable-driven design with 24 degrees of freedom and grip force over 30N, and a direct-drive design also in development; the motion-control algorithms originate from research at the University of Science and Technology of China, currently piloted in select industrial settings, with the company saying it aims to establish 1-2 flagship industrial projects by 2026. A high-reliability dexterous hand and a high-precision force-controlled arm are also in development, currently in testing or sample-shipping stages. source
· STMicroelectronics｜Accelerating robot chips for the South Korean market: the company says it can achieve a price advantage within 1-2 years. source
Top comments (0)