Today's Highlights
· Moore Threads (Chinese GPU maker) wins a 100,000-card order from JD Cloud for Chinese-made GPU clusters
· Pony.ai launches Robotaxi rides without safety drivers in Doha, which the company calls its first large-scale commercial deployment outside China
· Analog Devices acquires edge AI chipmaker Alif Semiconductor for $1.35 billion in cash, plus up to $200 million in contingent consideration
· Zhiyuan Robotics (Chinese humanoid startup) expands GE-Act 2.0's joint training data from 300 to 30,000 hours, lifting real-robot zero-shot average success rate from 17.1% to 44.1%
· PHYMI, the new company founded by former DeepRoute.ai (Chinese autonomous driving firm) vice president Liu Nianqiu, raises a seed round of nearly $100 million, led by IDG Capital with participation from Didi and Hesai
· Austin municipal records show 47 incidents of autonomous vehicles obstructing emergency response in Q2, as NHTSA simultaneously audits Cybercab's self-certification of compliance
Paper Progress
GE-Act 2.0: World-action model pretrained from scratch — zero-shot success rate doubles as data scales a hundredfold · world-model
Most world-action models to date have inherited off-the-shelf video generators, leaving open the question of how pretraining itself should be done and whether it scales. GE-Act 2.0 initializes both its generation and action components from random weights, trains solely on manipulation data, and uses KASO to select only predicted futures consistent with recorded action behavior for supervision. Evaluation uses the pretrained weights directly with no task-specific fine-tuning, across 100 tasks and 20 skill categories, with scene backgrounds, lighting, and object instances held out of the training set entirely: after scaling joint training data from 300 to 30,000 hours, G1-OP's average success rate rose from 17.1% to 44.1%, and G2-90D — which accounts for under 2% of the training data — also gained 17.7 percentage points, pointing to cross-embodiment transfer. The correlation between skill coverage and zero-shot OOD success rate reached a Pearson r of 0.80.
AgiBot Research Team (Zhiyuan Robotics) · arXiv 2609.05588 source · HF↑45
OpenWAM: Breaking world-action model pretraining into controlled experiments · world-model
Existing systems fuse the generative backbone, visual representation, architecture, information flow, inference process, and training data into a single block, making it impossible to ask which design choices actually matter. OpenWAM decomposes this design space into composable modules, paired with unified training, inference, deployment, and evaluation, then answers three questions through controlled experiments and releases OpenWAM-α: pretrained on roughly 6,400 hours of first-person human and robot data, it stays in the top tier across 8 simulation benchmarks and real-robot experiments, from single-arm to bimanual to dexterous-hand setups. Infrastructure, evaluation protocol, pretrained weights, and data recipes are all released.
Yuran Wang et al. · arXiv 2609.07398 source · HF↑47
TANGO: A humanoid navigates cluttered spaces by directly outputting 29-DoF joint actions · locomotion
Treating navigation as 2D path planning breaks down for a humanoid moving through a cluttered room, where arm positioning, torso lean, and gait all need continuous adjustment. TANGO takes natural language instructions and first-person RGB input and directly predicts 29-degree-of-freedom joint-space actions for a whole-body controller, with training done entirely in simulation using global path planning combined with kinematic generation, obstacle-avoidance editing, and RL tracking to synthesize feasible action supervision. Deployed zero-shot on a Unitree G1, it navigates cluttered real-world scenes following language instructions without any real navigation data.
Anqi Li et al. · arXiv 2609.09158 source · HF↑19
SimpleMemVLA: Skips a dedicated memory module, still sets new highs on four memory benchmarks · vla
Memory mechanisms like retrieval stores, learned compressors, and recurrent states all have to decide what to keep before knowing what will be needed later — a design premised on minute-scale history being too large to process directly, an assumption modern VLM backbones have already invalidated. SimpleMemVLA drops the dedicated memory module entirely, feeding sampled history straight into the backbone as timestamped video, and lets the generated subtask hidden state serve as the sole channel from history to the action head; adjacent decisions share most of their history, and common prefixes are pre-filled during action execution, keeping latency close to that of a single-frame VLA. It substantially outperforms retrieval, compression, and recurrent-state approaches across four memory benchmarks with no loss in general control capability, and causal intervention confirms the policy genuinely reads from history.
Cheng Yin et al. · arXiv 2609.05533 source
ContextFlow: Swapping in a flow-matching backbone for in-context imitation · manipulation
Continuing the in-context learning line of work: autoregressive in-context imitation discretizes continuous actions, and early prediction errors compound along the next-token chain, causing failure on unseen task configurations. ContextFlow feeds demonstrations and observations as conditioning into flow matching, using a perceiver-style multimodal compressor to distill vision, proprioception, and action sequences into a compact latent representation. On unseen LIBERO task configurations, it beats ICRT by 35 percentage points in average success rate, matching task-specific fine-tuned π0 with no fine-tuning at all; on a real robot, it achieves 40% on an unseen pen-cap removal configuration.
Jian Ding et al. · arXiv 2609.06852 source
GLoRI: Whole-body tracking corrects drift in the world coordinate frame · locomotion
Local references preserve motion structure but impose no constraint on absolute spatial position, letting global error accumulate. GLoRI uses global-local cross-attention to correct local keypoint features using global goal and pose-difference features. It achieves 100% completion rate and a g-MPJPE of 6.44 cm on held-out HuMoTo motions, transfers directly from Isaac Gym to MuJoCo without fine-tuning, and lets a single policy enable a real Unitree G1 to autonomously complete loco-manipulation with a variety of unfamiliar objects — whereas prior comparable systems mostly relied on teleoperation or handled only single objects.
Qingyao Xu et al. · arXiv 2609.05994 source
CosmoH2G: Human-hand-to-gripper transfer covers flips and rotations for the first time · manipulation
Transferring human hand demonstrations to robot grippers is a cheap data source, but existing methods are largely confined to simple planar tasks and fail on complex spatial trajectories involving rotation and flipping. The authors built a scalable hand-gripper paired data collection pipeline, producing a dataset of 6,189 episodes across 1,254 objects with significantly higher spatial complexity than existing benchmarks. The method proceeds in two stages: first predicting sparse start and end keyframes to simplify the mapping target, then generating the full continuous action based on them, while preserving learned orientation, applying a grasp heuristic to translation, and using kinematic-consistency post-optimization to suppress cumulative drift.
Hongxiang Zhao et al. · arXiv 2609.07498 source · HF↑25
SceneMosaic: Simulation-ready indoor scenes, generated 24x faster · benchmark
Agentic text-to-3D scene generation offers high fidelity but requires repeated placement and refinement, while parametric image-to-3D generation is fast but often produces physically implausible layouts — and both approaches struggle to produce diverse scenes from the same input. SceneMosaic first uses image priors to obtain initial candidates, then hands them to a VLM agent for evolution, exploiting the locality of natural scenes to decompose a scene into independent units that evolve separately before being combined via Cartesian product. On SceneEval-100, it matches the strongest agentic baseline in semantic layout quality while running 24x faster, substantially reduces physical violations, and scores highest in human evaluation; the code is open-sourced.
Xingjian Ran et al. · arXiv 2609.05594 source · HF↑41
Other papers today: Rethinking Safety for Generalist Robots — Sinha, Majumdar, Bajcsy, and colleagues argue that safety for generalist robots must go beyond collision and force to include context, user intent, and hard-to-model physical consequences (arXiv 2609.06326 source); WM-Craftnet treats the world model as recurrent task context rather than for imagination, making in-hand rotation more robust to pose shift and force disturbance (arXiv 2609.07002 source); MEMOBench uses process-level metrics to separate "forgetting" from "manipulation failure" (arXiv 2609.07047 source); Benchmarking Dexterity of Multifingered Robot Hands, from the US NSF HAND Engineering Research Center, presents a three-tier dexterity benchmark framework (arXiv 2609.05585 source); WHIRL converts binary human interventions into forward-looking risk signals for RL in real-world dexterous manipulation (arXiv 2609.06009 source); Large Discrete Policy abandons continuous denoising in favor of selecting from a large-scale vocabulary of physically plausible actions (arXiv 2609.07049 source); ARC-Bench shows that closed-loop replanning masks failures in action ordering within frozen JEPA world models (arXiv 2609.05461 source); LANTERN separates warning onset, hazard onset, and termination to measure the individual contribution of cooperative-driving warnings (arXiv 2609.06368 source); FIRE3D turns a single image or casually shot video into a simulation-ready 3D asset in under a minute (arXiv 2609.08848 source).
Open Source · Tools · Benchmarks
· First batch of JD embodied data released: The first batch from JD Cloud's initiative to collect 10 million hours of embodied data has been opened externally, drawing applications from institutions and universities in over 8 countries and around a hundred Chinese universities and research institutes; JD simultaneously released JoyAI-Echo WM, a real-time interactive world model source
· Hyper3D WorldGen: Generates editable 3D scenes from a single image, aimed at content production and simulation assets source
· General Robotics GRID: Claimed to be the first robotics intelligence platform to automate the full development-to-deployment lifecycle; the exact scope is defined by the vendor source
Funding and Deals
Analog Devices acquires Alif Semiconductor | All-cash | $1.35 billion | Plus up to $200 million in contingent consideration · hardware
ADI and Alif have signed a definitive agreement, approved by both companies' boards, with the deal expected to close before the end of calendar year 2026, pending expiration of the HSR antitrust waiting period. Alif makes AI-native microcontrollers and fusion processors, whose heterogeneous architecture supports real-time sensor fusion, low-latency inference, and on-device AI; its chips are already shipping in volume and have won design wins with consumer and industrial customers. ADI calls this line Physical Intelligence, meaning systems that can reason from signals such as motion, sound, vibration, radio waves, and thermodynamics, and run locally within power, latency, safety, and reliability constraints; ADI's fiscal year 2025 revenue exceeded $11 billion. Source: PR Newswire source
PHYMI | Seed round | Nearly $100 million · embodied
Led by IDG Capital, with participation from Yunqi Partners, Fosun RZ Capital, Yaotu Capital, as well as Didi and Hesai. Founder and CEO Liu Nianqiu started the company in early 2026 after more than seven years at DeepRoute.ai (Chinese autonomous driving startup), where he rose to vice president and partner and helped drive the deployment of the company's map-free autonomous driving system and its shift toward end-to-end and VLA architectures; earlier in his career he was a senior engineer at Intel. The company is pursuing what it calls "Physical Agent," combining cognitive reasoning with physical action to perceive, navigate, and manipulate in open, dynamic environments, with funding directed toward core model R&D, real-world dynamic data systems, and product engineering. Round classification differs across sources: the company's WeChat announcement and DealStreetAsia both describe it as a seed round, while 36Kr's English edition calls it a Series A. Source: DealStreetAsia source
Antioch | Series A | $32 million · world-model
Led by Greylock, with participation from A*, Category Ventures, Box Group, and Icehouse Ventures; angel investors include Palantir CTO Shyam Sankar, Foxglove CEO Adrian Macneil, and Nvidia executive Ian Andrews. The company builds a simulation testing platform for physical AI, calibrating simulations to customer hardware and running them at scale in the cloud, aiming to let robotics and industrial equipment changes be validated without always testing on real hardware. It closed an $8.5 million seed round earlier this year, currently counts Amazon's Ring among its customers, and partners with Nvidia and cloud provider Nebius. Source: AI Insider source
Zongwei Technology | Pre-C++ round | RMB 200 million · hardware
Led by Matrix Partners China, with existing shareholders Shunwei Capital, Lanchi Ventures, and Shengbao Capital participating; earlier investors include BYD, Huaye Tiancheng, and Innovation Works. Founded in Suzhou in 2020, this company's core business is intelligent magnetic-drive conveying; it has delivered over 2,000 projects to date with cumulative conveyor length exceeding 90,000 meters, covering 16-plus countries and regions, and the company says its 2025 order volume grew over 400% year-on-year. Its product line has extended into joint modules and frameless torque motors, with joint module rated torque ranging from 1.5 to 120 Nm (peak up to 450 Nm), and the D96*H50 series reaching a torque density of 81 Nm/kg. Source: Robotics Outlook source
Sinian Zhijia | Strategic investment | Amount undisclosed · autonomy
Yueda Capital invested, less than two months after the company's previous round. Sinian Zhijia builds autonomous heavy trucks for closed environments like ports; this round involved industrial capital rather than a pure financial investor. Source: Autohome source
Runway acquires Kinetix | Amount undisclosed · world-model
Video generation company Runway has acquired Kinetix, a Paris-based AI company, extending its world model capabilities toward robotics. Source: BlockBeats source
Estun completes 100% acquisition of Estun Cool Dexterous · industrial
Estun has bought out the remaining equity in its subsidiary Cool Dexterous, bringing the collaborative robotics business fully back under the listed parent company. Source: Sina Finance source
56 embodied robotics funding deals in China in August; top 5 took roughly RMB 14 billion combined · adjacent ⚠️ aggregated figures
According to Gasgoo Embodied Intelligence's tally, there were 56 funding deals in August across China's embodied robotics sector and core components. The single largest was XPeng's robotics business, which raised over $900 million, followed by Sharpa, which disclosed for the first time cumulative funding exceeding RMB 4.5 billion and a post-money valuation of roughly RMB 22 billion; Paxini Perception, Fifth Era Sciences, and Yuanli Wuxian each raised in the billion-RMB range, with the five combined totaling roughly RMB 14 billion. Of the 56 deals, nearly half were still at seed, angel, or pre-A stage, with relatively fewer growth-stage projects around Series B. The same tally also carries a cold data point: as of late August, Beijing's first humanoid robot data training center was reported to have ceased operations. The center was unveiled in March 2025 at Shougang Park in Shijingshan, spanned 3,000 square meters, and once deployed over 100 robots — going from launch to shutdown in under a year and a half. Partner Realman Intelligent's explanation is that the data-collection model shifted from lab-based teleoperation to remote teleoperation in real working scenarios. Source: Gasgoo Auto source
Commercialization and Deployment
Moore Threads wins 100,000-card order from JD Cloud; JD unveils 3-million-robot procurement plan same day · industrial ⚠️ plan figures
Moore Threads (Chinese GPU maker) announced it won an order for a 100,000-card GPU cluster from JD Cloud, to be used in JD Cloud's planned Chinese-made ultra-large-scale intelligent computing infrastructure, focused on large model training/inference and embodied intelligence; Moore Threads founder Zhang Jianzhong called it a key milestone toward the large-scale commercial use of Chinese-made intelligent computing. The same day, JD Logistics unveiled its "Wolf Pack" line of industrial robots at the JDDiscovery conference in Beijing: the "Smart Wolf" transport robot navigates between shelves, and the embodied dexterous arm "Yilang" completes package grasping, flipping, and placement within 10 seconds, alongside a dedicated model built to operate at -20°C, automated pharmacy sorting equipment, and delivery drones. JD Group Technology Committee Chairman and JD Cloud President Cao Peng laid out a five-year target of procuring 3 million robots, 1 million unmanned vehicles, and 100,000 drones, alongside more than 80 RoboBase robotics industrial bases, the first of which has broken ground in Guangzhou — these are plans, not orders in hand. JD's ecosystem currently employs roughly 700,000 delivery and logistics workers; founder Liu Qiangdong has previously proposed a "Nirvana Plan" to retrain frontline staff into robot maintenance and after-sales roles. On the data side, JD aims to collect over 10 million hours of real-world scenario video within two years; SCMP cited a June report from Shenwan Hongyuan estimating that the global stock of high-quality embodied interaction data available in early 2026 was around 500,000 hours — an order of magnitude short of the tens of millions of hours needed. Source: Xinhua source, South China Morning Post source, Guangzhou Daily source
Pony.ai launches Robotaxi without safety driver in Doha, ahead of its own testing-phase timeline · autonomy
The second AEMOB Forum, hosted by Qatar's Ministry of Transport, ran September 7-9 in Doha, where Pony.ai and Qatar's national transport company Mowasalat opened fully driverless rides to the public for the first time — the first appearance of autonomous rides without an in-vehicle safety driver in Doha. The two companies began public-road testing in August 2025 and moved to commercial operation just 4 months later; riders hail vehicles through Mowasalat's Karwa app, with service covering the historic old town, the central business district, tourist attractions, and trips to and from Hamad International Airport. The deployment uses Pony.ai's seventh-generation Robotaxi built for international markets, which the company calls its first large-scale commercial autonomous driving deployment outside China. Qatar's Ministry of Transport said in an official July statement that the project exceeded expectations on passenger satisfaction and safety performance, and that its technology and operations systems are ready for broader deployment. Pony.ai currently operates in 9 countries across the Middle East, Asia, and Europe, with public-facing paid rides already live in Croatia, Singapore, and Qatar, and has agreed with platforms including Uber to deploy over 4,000 vehicles outside China. Source: Gasgoo source
Lyft integrates Waymo rides into its app in Nashville · autonomy
Users in Nashville can now hail Waymo's driverless vehicles directly through the Lyft app. The two companies are combining their respective fleets and access points, following the same distribution approach Waymo has used with Uber in other cities. Source: Bloomberg source
Hai Robotics to deploy 1,500 climbing robots at a European fulfillment center · industrial
A single-site deployment of 1,500 tote-handling warehouse robots ranks among the largest orders in the European market. Source: Robotics & Automation News source
Vbot books over 5,000 quadruped orders at IFA; global cumulative sales top 13,000 units · adjacent ⚠️ vendor figures
During IFA 2026 in Berlin, Vbot signed procurement agreements with 16 international partners, adding over 5,000 new quadruped robot orders across Germany, France, and the UK. The product is the consumer-grade robot dog Vbot SuperDog (nicknamed Bubble), already in mass production and on sale, using lidar paired with UWB for autonomous following and obstacle avoidance, with a maximum payload of 12 kg and the ability to pull a small camping trailer. When pre-orders opened in December 2025, it received 6,540 orders worth roughly RMB 100 million; as of September 8, Vbot's quadruped robots have sold over 13,000 units cumulatively worldwide, across 15 countries and regions. These figures come from company press materials, and there is no third-party accounting of the order-to-delivery conversion rate. Source: GlobeNewswire source
UBTECH wins RMB 150 million bid for Leshan commercial service humanoid robot industrial project · humanoid
The winning bid amount is RMB 150 million, for a project positioned around commercial-service humanoid robot industrialization. Following the Shanxi coal mine project, UBTECH's order pipeline continues to skew toward local-government-led industrial projects. Source: Sina Finance source
TCS opens India's first "lights-out factory" lab in Pune · industrial
Tata Consultancy Services has set up this lab at its Sahyadri Park campus in Pune, called the Industrial Autonomy & Engineering Lab, centered on a fully robotic battery pack assembly line that lets manufacturers simulate and validate AI-driven autonomous production systems before deploying them on real production lines. It brings together digital twins, industrial AI, robotics, factory control systems, and real-time operational intelligence, covering use cases such as predictive maintenance, automated quality inspection, real-time process optimization, and human-robot collaboration. Sreenivasa Chakravarti, TCS's global head of industrial autonomy and engineering, said the lab provides a hands-on demonstration of how enterprises can move from digitized operations toward progressively autonomous production systems. Source: ET Manufacturing source
Industry Developments
The Information reports China's securities regulator has raised the IPO bar for humanoid robot companies · humanoid ⚠️ unconfirmed report
The report says China's securities regulator has given informal window guidance to some banks and companies that humanoid robot firms must show recurring revenue, narrowing losses, or genuine innovation to be considered for approval. As Reuters noted in its own coverage, Chinese financial regulators did not respond to requests for comment, and Reuters could not independently verify the report. Cited triggers include overheated early-stage fundraising this year, a lengthening queue of companies awaiting IPO, and recent declines in listed companies' share prices; Unitree Robotics, whose stock surged over fivefold on its trading debut, has since fallen back roughly 45%. Source: The Business Times source
47 autonomous-vehicle emergency-response obstructions in Austin in Q2; NHTSA examines Tesla's self-certification basis · autonomy
Austin municipal records show 47 incidents of autonomous vehicles obstructing emergency responders between April and June, with 16 occurring in June alone. Following the case opening reported last week, the audit inquiry is now formally numbered AQ26002, opened September 3 — the same day Tesla began charging for Cybercab rides — and disclosed the next day, covering roughly 1,000 vehicles, from NHTSA's Office of Defects Investigation. The review examines what technical data and processes Tesla is using to self-certify that Cybercab meets all Federal Motor Vehicle Safety Standards, and whether that self-certification rests on the position that "certain standards do not apply to vehicles without manual controls." NHTSA Administrator Jonathan Morrison said the agency supports the safe development of autonomous vehicles but that "as the federal regulator, we need to ensure all laws are being followed." The audit is not a recall and does not assess how well the vehicles drive on their own; no injuries or deaths have been publicly attributed to Cybercab to date. Tesla has chosen the self-certification path, while Zoox pursued a formal exemption, approved in late July with an annual cap of 2,500 vehicles. At the March 1 shooting scene on West 6th Street, a Waymo vehicle stopped in the roadway; Austin-Travis County EMS spokesperson Christa Stedman told Axios the vehicle did briefly block an ambulance's path, and the city later said as many as 5 autonomous vehicles froze near the scene, though EMS said it did not affect patient outcomes. Source: Medical Daily source
Zhiyuan releases two models in one day; AGILE 2.0 merges perception and motion control into a single system · humanoid
Alongside the GE-Act 2.0 paper, Zhiyuan Robotics also released AGILE 2.0 on September 9, a combined perception-and-control model using a vision-based end-to-end approach that integrates environmental perception, terrain understanding, whole-body motion control, and manipulation into one system. Equipped with this model, the Lingxi X2 humanoid has demonstrated ball-kicking while walking, long-rope jumping, and collaborative box-carrying, used to validate the robot's ability to sense its environment and adjust its actions in real time while moving. These are dynamic-balance demonstrations, distinct from sustained operation at production-line pace. Source: Shanghai Observer source, Guandian.cn source
Unitree says a world-action model drove fully autonomous sparring; the 73.6% figure in its IPO filing remains the main point of skepticism · humanoid ⚠️ vendor claims
In a video released the evening of September 7, Unitree's world-action model UnifoLM-X2-1.0 drove a humanoid robot through punches, blocks, dodges, and continuous attack-defense transitions; the company said the entire sequence relied on no external remote control or preset scripting, and said it validates the basic feasibility of large-scale deployment of world-model-driven humanoid robots. The most widely circulated point of skepticism traces back to Unitree's response to an IPO filing inquiry letter: in the first three quarters of 2025, scientific research and education accounted for 73.6% of Unitree's humanoid robot revenue, commercial/consumer use for 17.39%, and industrial scenarios such as smart manufacturing, inspection, and logistics combined for 9.01%. Wu Xiang, director of the Institute of Perception and Control at Zhejiang University of Technology, said the G1 was chosen because few mature robot platforms were available at the time and Unitree offered the most stable, cost-effective option, with the EDU version supporting further development, allowing new motions to be trained in simulation and then deployed and validated on real hardware. Huang Siyuan, director of the Embodied Robotics Center at the Beijing Institute for General Artificial Intelligence, said Boston Dynamics' Spot sells for roughly $75,000 with very limited shipments, and its Atlas was not sold externally in its early phase, while Unitree drove prices down while still maintaining shipment volume. A sole-source procurement notice from Westlake University stated that leading foreign products are expensive and subject to technology controls against China, while other Chinese products still fall short on perception and motion capability. Wang Xingxing said publicly that day that he hopes to genuinely achieve robot self-evolution within the next year or two.Source: Hangzhou.com.cn source, China News Service source
NEURA teams up with SECO and Qualcomm to bring humanoid compute modules to mass production · humanoid
The division of labor: NEURA contributes physical AI and robotics architecture, Qualcomm contributes the edge computing platform, and SECO is responsible for engineering Qualcomm's Dragonwing processors into a mass-producible compute module. NEURA's robots are built around a "Brain + Nervous System" architecture, with its Smart Limbs distributing perception and compute close to joints and sensors rather than centralizing them. The two companies will also jointly collect data from industrial deployments to build physical AI automation solutions for semiconductor and electronics manufacturing. Founder and CEO David Reger said physical AI won't be built by a single company — it requires an ecosystem spanning sensing, compute, cloud, and industrialization. The partnership also includes establishing a NEURA Gym in Italy, the company's first training center in southern Europe. Source: Evertiq source
Samsung Medical Center to demonstrate two humanoid surgical assistant robots working together on the 16th · embodied ⚠️ planned demo
This hospital leads South Korea's Ministry of Health and Welfare's Korean ARPA-H program, on a project titled "Development of Humanoid Physical AI Surgical Assistant Robots for Efficient Surgical Environments," with a consortium called ORchestra that includes Rainbow Robotics, Aidin Robotics, Hahaeho, Samsung Advanced Institute for Health Sciences and Technology, Seoul National University, the National Cancer Center, and Jeonbuk National University Hospital, with about a year of development so far. The demonstration is planned for the 16th: Robot No. 1 will act as a surgical nurse, while Robot No. 2 will simultaneously handle both assistant-surgeon and nurse tasks; in the first demonstration, the robot itself holds surgical instruments and performs continuous actions such as incision, dissection, and suturing on biological tissue, with the operator watching a monitor and the robot replicating the operator's arm movements. Professor Jeong Yong-gi said, "We're not yet at the stage where the robot can judge the patient's condition on its own and perform surgery automatically." Development is currently at the stage of validating instrument delivery/retrieval, endoscope operation, tissue retraction, voice-command understanding, and safe stop-and-resume within a simulated surgical environment; commercialization would require completing performance testing, simulation and preclinical trials, clinical usability assessment by medical staff, GMP quality systems, clinical trials, and medical device approval. Source: Seoul Economic Daily source
NIO's Li Bin draws the line: embodied intelligence stays a long-term strategic investment, not part of NIO's core business · autonomy
NIO founder Li Bin said the company will engage with embodied intelligence through strategic investment for the foreseeable future rather than developing it into a core business line. He also called for the company to focus on its core business over the next three years. Last month NIO made a strategic investment in a new company founded by Ren Shaoqing.Source: TMTPost source, CnEVPost source
Hardware · Supply Chain
· National standard for dexterous hands: China's national standard for dexterous hands is expected to be published for public comment by year-end source
· Axera M9 series: A dual-chip solution for advanced driver assistance has officially launched source
· Zhaowei Machinery & Electronics: General manager of the dexterous hand business Chen Yidong said dexterous hand iteration has been extremely fast over the past three years with heavy capital investment, and that whether robots are worth deploying comes down to output ratio and cycle time source
· DFI Mini-ITX platform: A single motherboard compatible with two CPU generations, aimed at physical AI applications in robotics and automation source
Top comments (0)