Today's Highlights
· XPeng's first IRON humanoid drove itself off the production line, with core manufacturing automation exceeding 80%
· Inceptio Technology's (Chinese autonomous trucking company) commercialized autonomous driving mileage surpassed 1 billion km, alongside the launch of a freight-focused physical AI framework
· UBTECH (Chinese humanoid robotics company) sold 921 units of full-size humanoids in H1, accounting for nearly half its revenue
· Motional open-sourced nuReasoning, with 20,000 long-tail scenarios paired with human-verified reasoning annotations
· Li Auto's MindVLA-o1 brought 3D ViT into the vehicle via its self-developed 1280 TOPS chip
· Arm introduced the Robotics Capability Framework, aiming to establish a unified vocabulary for robot capabilities
Research Progress
HSImul3R: Making reconstructed people actually sit properly on chairs in a simulator · perception
3D reconstruction has long optimized only for "does it look right" — a human body visually accurately seated on a chair can, once dropped into a physics simulator, tip the chair over or clip through it. Dexmal (Chinese robotics company) partnered with Nanyang Technological University's S-Lab and Shanghai AI Laboratory to write gravitational stability, physical contact realism, and interaction stability into the evaluation criteria, using a closed-loop physical optimization: scene-oriented reinforcement learning corrects human motion on one side, while direct simulation-reward optimization feeds back to correct 3D object generation on the other. On their self-built HSIBench, interaction stability rates across Easy/Medium/Hard tiers were 53.68%, 30.56%, and 13.92%, versus HSfM's 10.52%, 4.50%, and 2.66%; the human-scene interpenetration rate dropped from 69.51% to 22.90%. The optimized motions were ultimately transferred to a Unitree G1 for execution, converting everyday video interaction experience into skills that are both simulatable and learnable by robots.
Dexmal × Nanyang Technological University S-Lab × Shanghai AI Laboratory · Accepted to ECCV 2026 · Reported by: GeekPark source
WM-LOCO: Training a world model and a walking policy together to cross a 0.8-meter gap · locomotion
Conventional approaches to walking on complex terrain require labeled footholds, teacher distillation, and multiple chained modules. Geek+ — wait, this is DEEP Robotics — placed a world model for future visual and motion-state prediction and a PPO policy into the same end-to-end training process, going directly from depth visual input to joint commands. The same set of weights was deployed zero-shot to a Unitree G1; across three terrain types — stepping stones, stairs, and gaps — tested 10 times each, the average success rate was 93.3%; an 0.8-meter-wide gap was cleared in a single complete swing-cross-land motion, and the team stopped there for safety reasons rather than widening it further. The entire algorithm runs in real-time closed loop on the Horizon Sunrise S600 compute platform.
DEEP Robotics · Reported by: Gasgoo source
HERON World Model: Cutting general internet video out of the training set entirely · world-model
Where the industry's standard data pyramid is built on a foundation of massive internet video, AGIBOT's (Chinese humanoid robotics company) new sub-brand Scalabot removed that layer entirely, keeping only three types of data: real robot teleoperation data, simulation data rebuilt from real-world scenes, and first-person human operation video. The reasoning isn't that there isn't enough data overall, but that there isn't enough of the right data: world models need contact-dense video, and that category is a vanishingly small share of internet video — filtering for it costs more than direct collection. Heterogeneous actions aren't forced into a unified format; instead, physical-space alignment is done first, then each domain's lightweight encoder projects them into fixed-dimension Action Tokens. Evaluated offline under the open test protocol of WorldArena 1.0 Track 1, the composite score EWMScore_P was 75.80.
AGIBOT Scalabot · Reported by: Robot Lecture Hall source
Open Source · Tools · Benchmarks
· nuReasoning (Motional): the first reasoning-centric open-source autonomous driving dataset, cutting 105 hours of real road-test footage from fleets in Las Vegas, Pittsburgh, Los Angeles, Boston, and Singapore into 20,000 long-tail scenarios, each at least 20 seconds of video, paired with 247,000 human-verified reasoning annotations, used to train VLA models to understand the causal reasoning behind driving decisions rather than merely imitating trajectories. A companion challenge launched alongside it, split into explainable motion planning and long-tail scenario reasoning tracks, with results to be announced at NeurIPS in December; the miniset released earlier this year has already logged over 50,000 downloads. This line traces back through nuScenes (2019), nuImages, and nuPlan. source
Funding & Deals
Qianjue Technology (Chinese world-model startup) | Series A+ | Several hundred million RMB · world-model
Jointly invested by Vision Plus Capital, Yuanhe Chinese company's fund Houwang, InnoAngel Fund, Xiaoguang Capital, Xinneng Venture Capital, and Jingming Capital, among others. The team spun out of Tsinghua University's Brain-Inspired Computing Research Center dual-arm robotics group, with the group's former lead, Dr. Gao Haichuan, now CEO. The technical path benchmarks against Yann LeCun's team's JEPA, building predictive world models and polynomial representation architectures; last year's Causal Dreamer, which introduced a counterfactual filtering mechanism, was published in Neurocomputing. The company says it has completed intelligent adaptation for nearly 50 robot models, with over 100,000 units deployed in the field, and holds 2 billion real-scene 3D perception data assets. ⚠️ Vendor-stated figures Founded in 2023, this is its 9th funding round.Source: ThinkinChina (Zhidongxi) · Robot Outlook source
Xiaoyu Zhizao (Chinese welding-robot startup) | Series B3 | Several hundred million RMB · embodied
Funded by Blooming Star, Singapore's Kamet Capital, and Guoke Yingfeng Fund, closing in August — the company's third round in 2026. Founder Qiao Zhongliang came from Xiaomi's core founding team; the technical premise is "one brain, many forms," a single general-purpose brain adapted to different physical forms for welding, polishing, precision assembly, and inspection. The company chose intelligent welding as its entry point; the standard version of its Xiaoyu Future Robot smart welding workstation, launched in July, starts at RMB 169,800, with deployment scenarios extending from architectural steel structures to shipbuilding and heavy machinery. The company says roughly 100,000 real production data points are transmitted back in real time from customer sites daily, with over 100,000 cumulative hours of data collected.Source: PEdaily source
Tianji Intelligent (Chinese force-control robotics company) | Series B++ strategic round | Amount undisclosed · hardware
Co-led by Ant Group and SOFINA, with existing investors Banyan Capital and Tencent, among others, participating. This Guangdong-based company's core competency is force control: MEMS joint torque sensors, lightweight integrated joint modules, force-control algorithms, and control systems, all the way up to force-controlled humanoid arms. Its newly incubated humanoid brand Gento has begun deliveries: Luna is a wheeled, foldable model with 26 degrees of freedom, standing 1.7 meters tall and folding down to 0.79 meters — small enough to fit in a standard elevator; Skye is a wheeled, height-adjustable model with 23 degrees of freedom, a lift range of 785 to 1695 mm, designed for continuous 7×24 operation and dedicated to reinforcement learning training and real-robot data collection.Source: PEdaily source
AgiBot (Chinese embodied-AI startup, distinct from AGIBOT/Scalabot above) | Confidential filing with Hong Kong Stock Exchange | Over RMB 5 billion raised in two and a half years, valuation of RMB 20 billion · embodied ⚠️ Reported, unconfirmed
Over RMB 5 billion from more than 30 institutions has flowed into a company that has yet to achieve large-scale production and sales revenue, over two and a half years. Meituan, Alibaba, ByteDance, and Xiaomi have each led different rounds; Sequoia China has participated in every round since the Series A+, with the valuation jumping from RMB 10 billion to RMB 20 billion in just over two months. On August 5, the South China Morning Post cited two people familiar with the matter saying the company had confidentially filed for a Hong Kong listing. The August 12 livestream at an SF Express warehouse became the focal point of controversy: the official result claimed sorting of 1,816 randomly assigned parcels in one hour with a success rate above 98%; a blogger known as "Robot Detective" broke down the numbers to argue that Figure's comparable figure of 1,248 parcels/hour came from a 200-hour continuous livestream averaged across multiple rotating robots totaling roughly 249,000 parcels, and that Figure used full-size humanoids while AgiBot used fixed-station dual arms with standard grippers. Another point of contrast is the home setting: as recently as March, the company was selling three-hour home cleaning sessions for RMB 149 on the platform 58 Daojia; it has now shifted to offering residents up to RMB 3,000 per month to let robots into their homes to collect data.Source: Huxiu, republished via Ifeng Tech source
Wang Naiyan's Embodied Brain Project | Fundraising in progress | Amount undisclosed · embodied ⚠️ Single-source account
Wang Naiyan, former head of L3 autonomous driving technology at Xiaomi and former CTO of TuSimple's China operations, has launched a startup focused on embodied "brain" technology; one investor says the core R&D team numbers a dozen or so people, many of them former colleagues from his TuSimple days. The distinctive part of the approach is reusing L3 logic — placing reinforcement learning into the pretraining stage of the action model rather than treating it as a final fine-tuning step: during pretraining, domain randomization is applied, with only a handful of first-principles constraints — safety floors and physical laws — set as the reward function, letting the model self-play thousands upon thousands of times. "Actions aren't learned by copying imitation, they're learned by deduction," is how one person familiar with the matter paraphrased his core philosophy. According to IT Juzi, embodied-brain systems now account for 38.8% of funding round counts this year, overtaking humanoid robots' 21.1%, though roughly 70% of the more than RMB 46 billion disclosed in China in H1 flowed to the top 20 companies.Source: 21st Century Business Herald · Yijian Auto source
Commercialization & Deployment
XPeng activates humanoid production line; first IRON walks itself off the line after completing automated final assembly · humanoid
The first IRON unit to complete automated final assembly on this line wasn't carried off — it walked to the camera on its own. The line was designed in-house by XPeng, developed from the ground up with mass production as the goal, with core manufacturing automation exceeding 80%, bringing the automotive industry's quality system into precision robotics manufacturing and controlling mass-production quality to automotive-grade consistency standards. The latest IRON stands roughly 170 cm tall, with 76 degrees of freedom across the body and 21 in each hand, uses a fully enclosed flexible lattice covering, and carries 3 Turing AI chips delivering 2250 TOPS of effective compute; physical AI models run entirely on-device, allowing it to perform complex tasks autonomously without teleoperation. Per the roadmap, IRON will enter mass production by the end of this year, first deployed in XPeng stores and campuses, with delivery to markets in China and overseas planned for 2027. The over $900 million Series A round two weeks ago, at a post-money valuation exceeding $6.3 billion, was led by IDG Capital, with Tencent and Alibaba entering as strategic investors (previously reported). Producing one working demo unit and reliably replicating tens of thousands are not the same order of difficulty; the production line is a fact on the ground, while end-of-year mass production and 2027 delivery remain a schedule.Source: Guancha source
Inceptio Technology's commercialized autonomous driving mileage surpasses 1 billion km; launches Freight Physical AI · autonomy
1 billion km of commercialized autonomous driving mileage, with the system covering roughly 97% of China's expressway network. The company says it holds over 90% of China's autonomous trucking market. Freight Physical AI, launched the same day, is an intelligent framework for commercial freight vehicles, combining freight-specific training data, a scenario library built from operational records, and a cloud platform connecting vehicle-side and fleet-side intelligence. Several leading express delivery companies have already made autonomous driving standard on heavy-truck orders this year, which the company treats as a signal of the technology's commercial viability. On the licensing front, it holds L4 permits in Beijing, Zhejiang, and Hainan, has run open-road cargo demonstrations with JD Logistics, and in May put unmanned light trucks into a commercial pilot on open roads with SF Express, extending autonomous freight into regional and urban delivery. Overseas, it is running validation in Europe, Japan, and the Middle East, and has been selected for a pilot project at the Port of Antwerp-Bruges in Belgium.Source: Automotive World source
UBTECH sold 921 full-size humanoids in H1, accounting for 46.5% of revenue · humanoid
Following last week's H1 earnings reports from seven listed robotics companies, a September 7 research note from China Galaxy Securities broke down UBTECH's numbers: 921 units of full-size embodied-AI humanoids sold in H1 2026, up 1946.7% year-on-year, corresponding to revenue of RMB 590 million, up 1445.0% year-on-year, and accounting for 46.5% of total revenue. Overall operating revenue was RMB 1.269 billion, up 104.2% year-on-year; gross margin was 44.7%, up 9.7 percentage points year-on-year; net loss for the period was RMB 339 million, versus RMB 440 million in the same period last year; adjusted EBITDA was -RMB 174 million. All three major expense ratios declined; R&D spending rose to RMB 303 million, but the R&D expense ratio fell by 11.2 percentage points. On the logistics robotics side, the company completed project deliveries for Fortune 500 clients including BOE, Jabil, Foxconn, and Honda.Source: Sina Finance, citing a China Galaxy Securities research note (analysts Lu Pei, Peng Xingjia) source
Yuechuan Bionics' W-Bot 2.0 deployed at Hefei's State Grid site for live-line cable stripping · embodied
The wheeled humanoid W-Bot 2.0 completed a full end-to-end cable-stripping operation at a State Grid site in Hefei, in a live-scenario validation for uninterrupted urban distribution-grid maintenance. It was deployed mounted on an insulated bucket truck, equipped with a Xinshou (RealMan) X-Hand M1 bionic dexterous hand and a 7-DOF humanoid arm, with a payload of 10kg per arm and 20kg for both arms combined, covering an operating height range of 0 to 2.2 meters; its base footprint is only 0.2 square meters, small enough to maneuver through narrow gaps between wires. During live-line operation, a whole-body hierarchical neural-network balance controller computes center-of-mass shifts in real time, maintaining stance even under significantly increased single-arm loads to avoid contact with nearby live components. This is currently a single-site field validation, not routine operational maintenance.Source: Gasgoo source
Hai Robotics partners with SSI Schaefer on Australia's first climbing-robot warehouse project · industrial
The two companies have already jointly delivered over 30 projects in Europe and are now bringing that experience to the Australia-New Zealand market. Nathan Zeng, Hai Robotics' President for Asia Pacific, said the partnership with SSI Schaefer is one important strategy but not the only one; the company is also pursuing direct customer contact, local service capability, and a partner network in parallel. "What we design and manufacture is something SSI Schaefer doesn't have, so this isn't competition — it complements their product line." The tote-handling solution is positioned to avoid overlap with SSI Schaefer's existing automated storage-and-retrieval systems.Source: MHD Supply Chain source
Talabat and QuikBot pilot high-rise delivery robots in Dubai · adjacent
The two companies have signed a memorandum of understanding covering the segment of delivery that begins after a rider reaches the building: through the lobby, into the elevator, and to the resident's door. Singapore-based QuikBot's platform provides floor-to-floor delivery for apartments, offices, and hotels, along with smart lockers for handoff. The number of buildings and robots involved, when residents will gain access, and whether pricing will change during the pilot have not been disclosed.⚠️ Plan-stage announcement This is infrastructure testing within buildings only; the full restaurant-to-door journey will continue to be handled by Talabat's existing rider network.Source: tbreak source
Industry Developments
Li Auto unveils next-generation autonomous driving architecture MindVLA-o1, bringing native 3D ViT into the perception stack · autonomy
At NVIDIA GTC 2026, Li Auto's foundation model lead Zhan Kun offered a diagnosis: today's end-to-end systems are essentially "learning to drive by watching 2D video" — BEV flattens the world and loses height, while OCC has 3D structure but lacks semantics. Li Auto's solution is to introduce a native 3D ViT, achieving unified geometric and semantic understanding of 3D space directly at the encoding stage, with lidar's role shrinking from the core of perception to a high-precision ruler providing geometric calibration. Li Xiang says the unified modeling enables stable perception and reasoning out to over 500 meters. This approach was previously not viable on-vehicle due to insufficient onboard compute; the self-developed Mach 100 chip, delivering 1280 TOPS per unit, made deployment possible, first fitted to the all-new L9. The architecture is centered on a native multimodal MoE Transformer, alongside a predictive latent world model, unified behavior generation via VLA-MoE, discrete diffusion trajectory optimization, and closed-loop reinforcement learning; scene reconstruction has shifted to a feed-forward approach, roughly doubling rendering speed while cutting overall training cost by about 75%. The foundation model team evaluated nearly 2,000 model architecture configurations for hardware-software co-design.Source: Autohome source
Arm launches Robotics Capability Framework, aiming first to unify how the industry talks about capabilities · adjacent
Arm's Chief Architect Richard Grisenthwaite argues that terms like autonomous, intelligent, adaptive, collaborative, context-aware, and self-improving are everywhere today, but don't mean the same thing across different vendors, use cases, operating environments, and supervision models. The result is that systems are hard to compare, hard to integrate, and hard to evaluate: customers can't tell what capability they're actually buying, regulators and insurers lack a clear basis for assessing capability versus risk, and developers and integrators lack a shared vocabulary. Arm's Robotics Capability Framework is an architecture-agnostic reference framework describing capabilities, operating context, supervision modes, and assurance levels — it does not prescribe how a robot should be built. Grisenthwaite has said explicitly that this isn't the end point, nor is it intended to become something Arm alone owns. The same day, STRADVISION and ZaiNar announced they are bringing visual perception and localization capabilities into Arm Total Design for Physical AI.Source: Arm (as covered by GamesBeat) source
Samsung SDS sets KRW 10 trillion investment plan, with the US as its main robotics battleground · industrial
CEO Lee Joon-hee said at the REAL Summit 2026 at Coex in Seoul that the company will invest KRW 10 trillion (roughly $7.4 billion) by 2031, and that it is "evaluating investments spanning AI transformation and robotics, including taking a stake in a robotics foundation model developer." This sum builds on an April plan of the same size focused primarily on data centers and GPUs, now adding strategic investment and M&A, with funding sources including $820 million from KKR and KRW 6.6 trillion in cash and equivalents. The three pillars are AI transformation, robotics transformation, and logistics; for robotics, the company is treating the US as its primary market, citing the push there for manufacturing reshoring. ⚠️ Plan-stage announcement The company also says it will launch a robot orchestration platform in 2027 to manage mixed robot fleets across more than 1,000 production lines.Source: Korea JoongAng Daily source
Palladyne AI and FANUC America form strategic partnership · industrial
The partnership pairs FANUC America's industrial robot product line with Palladyne AI's physical AI software platform, targeting manufacturing, warehousing, and logistics. Six focus areas include optimizing Palladyne IQ on the FANUC platform, AI-driven motion planning and adaptive behavior, teleoperation and human-assisted learning, simulation and model training to accelerate deployment, joint validation of customer use cases across scenarios, and standardized deployment processes for system integrators. ⚠️ Single-party statement The announcement disclosed no financial terms, development timeline, or named initial customers.Source: Joint announcement by Palladyne AI and FANUC America source
Unitree claims UnifoLM-X2-1.0 achieves fully autonomous humanoid combat · humanoid ⚠️ Vendor-stated claim
In a video released September 7, Unitree says the model achieved a breakthrough in instantaneous planning, decision-making, and dynamic interaction execution within its action foundation model, enabling high dynamism, strong interaction, and real-time prediction and planning for future states, and says this validates the basic feasibility of large-scale deployment of world-model-driven humanoid robots. What was made public is a video, not reproducible results or mass-production specifications.Source: China.com source
Dreamer creator Danijar Hafner leaves DeepMind to start a company, office walls lined with humanoids imported from China · world-model
31-year-old Hafner's office in San Francisco's SoMa doesn't even have a nameplate on the door yet, furniture is sparse — but a row of various humanoid robots imported from China hangs from the shelves. His approach is model-based reinforcement learning: first build a world model that simulates physical reality, let the agent learn to act within it, then use that experience to predict future outcomes, enabling it to handle environments it has never encountered during training. Unlike traditional robot learning, which relies on real-world trial and error, this approach lets robots perform complex tasks in scenarios they've never run before. He became a student researcher at Google Brain in 2015, while still a college sophomore, and has since moved between the UK, Canada, and the US across stints at Google Brain and DeepMind. DeepMind's Timothy Lillicrap said: "I've met a lot of very smart people in Google Research, and he easily ranks in the top 0.5%."Source: MIT Technology Review source
Hardware · Supply Chain
· Axera M97 / M95: two 5nm automotive-grade autonomous-driving chips; the flagship M97 integrates 720 TOPS of effective AI compute, 16 Arm Cortex-A78AE cores, and a proprietary ISP supporting 16 camera inputs, with effective memory bandwidth of 460GB/s; the mainstream-tier M95 offers 360 TOPS, 10 cores, and 14 camera inputs. Both chips have passed AEC-Q100 automotive-grade certification, feature a built-in ASIL-D on-chip lockstep safety island, reach ASIL-B at the overall chip level, and support dual-chip redundancy. The company says end-to-end and VLA model performance significantly exceeds industry norms, with ISP pipeline latency reduced by 70%. source
· Korean battery material makers, including EcoPro, pivot toward humanoid robots: EcoPro plans to accelerate development of high-nickel cathodes and next-generation solid-state batteries to supply humanoid robot makers; LG Energy Solution, Samsung SDI, and SK On have previously announced similar plans. Bloomberg notes that humanoid robots typically require higher energy density than EV batteries, offering Korean manufacturers — who have been losing LFP battery market share to Chinese rivals — a potential opening. CRU battery materials analysis head Sam Adham said solid-state batteries could further boost energy density, and that the robotics industry has a higher cost tolerance than automakers, though he expects the industry won't truly take off until the latter half of the next decade. EcoPro currently operates a pilot plant with 40 tonnes annual capacity for sulfide-based solid electrolytes, with commercial production planned to begin next year. source
I caught myself mistranslating "地瓜机器人" as "Geek+" mid-sentence and self-corrected in the visible text — that shouldn't have been left in. Let me fix that.
WM-LOCO: Training a world model and a walking policy together to cross a 0.8-meter gap · locomotion
Conventional approaches to walking on complex terrain require labeled footholds, teacher distillation, and multiple chained modules. DEEP Robotics (Chinese quadruped/humanoid robotics company) placed a world model for future visual and motion-state prediction and a PPO policy into the same end-to-end training process, going directly from depth visual input to joint commands. The same set of weights was deployed zero-shot to a Unitree G1; across three terrain types — stepping stones, stairs, and gaps — tested 10 times each, the average success rate was 93.3%; an 0.8-meter-wide gap was cleared in a single complete swing-cross-land motion, and the team stopped there for safety reasons rather than widening it further. The entire algorithm runs in real-time closed loop on the Horizon Sunrise S600 compute platform.
DEEP Robotics · Reported by: Gasgoo source
Top comments (0)