Today's Highlights

· Tesla's Optimus enters production in Fremont; Musk says 20,000 units per week is at least a year away, and external deliveries won't begin until the end of 2027

· Tesla owners' personal cars won't be able to join the Robotaxi network until Q2 2027, and unsupervised FSD will only work inside the operating geofence

· Deepal SL03 (Changan EV model) receives the first dedicated L3 license plates; the feature can only be activated on three Chongqing expressway sections at speeds below 50km/h

· Agility names Nvidia the sole onboard compute supplier for Digit 5, with early supply set for H1 2027

· DONIC tracks bimanual dexterous manipulation with a single policy, reaching an 86.79% success rate on unseen trajectories, 77 percentage points above the strongest baseline

DONIC: One Policy for Tracking Bimanual Dexterous Manipulationmanipulation

HF 99↑, the highest community buzz today. Most earlier dexterous tracking work covered only a single hand, because data on sustained two-handed coordinated contact is scarce and the coordination is hard. DONIC puts PPO and behavior cloning in the same loop, recycling successfully executed trajectories as new demonstrations; the accompanying BiHOI dataset has 229 objects and 1,575 bimanual interaction clips. With a single checkpoint and no fine-tuning, success is 86.79% on unseen trajectories of training objects and 57.32% on new objects, which is 77.36 and 45.73 percentage points above the strongest bimanual tracking baseline, respectively.

Mingze Sun et al. · arXiv 2610.29997 source

COWBOY: A Scalable Whole-Body Sim2Real Framework for Mobile Manipulationmanipulation

A mobile manipulation robot has to decide each time where to park its base, where to point its camera, and how to grasp. COWBOY works in three steps: first, a vectorized whole-body controller solves "reachability" across thousands of scenes; then RL experts are trained in local scenes to handle grasping; finally the two are distilled into one vision policy that controls the base, arm, dexterous hand and movable camera directly from point clouds. The policy transfers zero-shot to real hardware and grasped hundreds of objects in cluttered scenes such as shelves and drawers.

Jiahui Yang et al. · arXiv 2610.28948 source

Real2Sem2Real: One Human Hand Video Yields Thousands of Robot Demonstrations on Similar Objectsmanipulation

The only input is a single monocular RGB video. The system tracks object pose with a DINOv3-distilled Gaussian splat, aligns "digital cousins" generated from one or two photos to the demonstrated object by functional region, then generates demonstrations in bulk in IsaacSim or MuJoCo, producing more than 2,000 in about two hours. Experiments cover two real robots and four tasks; for each task a VLA trained on just one human video matches or exceeds a policy trained on teleoperation data.

Matthew Strong et al. · arXiv 2610.28880 source

Six Frontier LLMs Directly Operate a Robot Arm on LIBERO-10benchmark

This week brings another evaluation of general-purpose LLMs controlling robots directly. The setup involves no fine-tuning and no pretrained skills: the model sees only images and proprioceptive state, and specifies end-effector pose and gripper opening/closing through tool calls. Six models were each run at three reasoning-effort levels, for 1,800 episodes in total. GPT-6 Astra had the highest success rate at 44%, with Gemini-4 Argon at 43%. Raising reasoning effort usually helps, but the gain varies by model and inference cost rises with it; placement-type tasks are handled reasonably well, while precise positioning and multi-step interaction remain poor. The evaluation framework and results are public.

May Espinola et al. · arXiv 2610.30051 source

CADRE: Rebuilding the Video Autoencoder for Interactive World Modelsworld-model

Most video world models borrow the autoencoder from a text-to-video model, with a fixed compression rate and no way to retrain it. Kyutai has released the CADRE family of causal diffusion autoencoders, covering multiple spatiotemporal compression rates, and compares them one by one using the same world model on manipulation and navigation tasks. One finding is that the two task types need opposite compression directions: for manipulation with a fixed camera, compressing the time dimension costs little, while for navigation with a moving camera, compressing the spatial dimension costs little. Code, models and an online demo are public.

Adrien Ramanana Rahary et al. (Kyutai) · arXiv 2610.30129 source

AW-0: Putting Code Planning into a World Action Modelworld-model

HF 8↑. AW-0 places code, future video and actions as three coupled generation streams in a single mixture-of-transformers. A code expert writes parameterized skill programs, a video expert generates future frames, and an action expert outputs motor commands. An accompanying auto-annotation pipeline labels robot video and human video with aligned programs, then organizes them into a reusable skill library. On RoboTwin 2.0 and real dual-arm tasks, AW-0 outperforms VLA and WAM baselines.

Zhuoyang Liu et al. · arXiv 2610.30136 source

SERA: A Self-Modifying Planning Layer for History-Dependent Manipulation Tasksvla

HF 9↑. A fixed VLA executes actions at the low level, while an upper-level VLM tracks task progress through an editable harness. The harness proposes and filters its own modifications using environment rollouts, and the selected version is then run on the test set. On RoboMME, SERA's overall success rate is 51.38%, versus 44.38% without self-evolution.

Tingwen Du et al. · arXiv 2610.29560 source

Watch Your Step: Looking Only at Terrain Near the Next Footsteplocomotion

Perceptive locomotion policies usually ingest an entire elevation map. This work first uses a lightweight model to predict the next footstep location from proprioception and velocity commands alone, then gives the policy only the small terrain patch around that footstep. The training iterations needed to reach the same terrain difficulty are about one third of the original, with performance comparable to an attention encoder that costs several times more to train. The policy transfers zero-shot to the Unitree G1, which can climb stairs, cross gaps and descend steps.

Matthias Heyrman et al. (ETH Zurich) · arXiv 2610.29612 source

Survey of Multimodal Embodied Agents: From Computer Use to Robot Manipulationbenchmark

HF 17↑. The survey proposes the PAPAV framework, which breaks agent capability into five parts: perception, anticipation, planning, action and verification. The authors surveyed 64 benchmarks; only 5 test anticipation on its own and 3 test verification, while the rest look only at task success rate.

Yanzhe Chen et al. · arXiv 2610.30108 source

Huixi Intelligence (Chinese embodied-AI chip developer) ｜ Pre-B+ and Series B ｜ More than RMB 2.5 billion raised this year ｜ Valuation undisclosedhardware

The Beijing Municipal AI Industry Investment Fund led, with participation from the Beijing Robot Industry Development Investment Fund, the National AI Industry Investment Fund, Yinshan Capital, Zhengxin Valley Capital and others; existing shareholders including Shunwei Capital (Chinese VC firm), Cathay Capital and SenseTime Guoxiang Capital doubled down. Huixi makes high-compute chips for embodied intelligence, and a computing platform based on its Guangzhi R1 SoC is installed on the AgiBot Genie G2. The company says that in June multiple G2 units livestreamed production-line work for 6 consecutive days, with 17625 cumulative operations at a 99.99% success rate. A Sohu funding flash puts the Series B amount at RMB 1.7 billion, while the company's press release gives only the cumulative figure of more than RMB 2.5 billion this year. Source: 36Kr source; Sohu source

Reliable Robotics (US) ｜ Series D extension ｜ $115 million added, $275 million total in Series D ｜ Valuation above $1 billionautonomy

The Series D was originally set at $160 million. Existing investor Nimble Partners led this extension, and new investors include Gideon Capital, RockCreek and Geek Ventures. The new portion includes a $25 million term loan from Silicon Valley Bank; the company's total equity funding exceeds $380 million. It develops RAS, an autopilot system for fixed-wing aircraft, with the goal of FAA certification on the Cessna 208B covering taxi, takeoff, en route flight and landing. This week it completed its first automated flights at Albuquerque International Sunport: a remote pilot in Mountain View, more than 700 miles away, supervised, with a safety pilot also on board. Source: AI Insider source

Enact (Switzerland) ｜ Pre-seed ｜ $30 million ｜ Valuation undisclosedembodied

Institutional investors are Redalpine, Notion Capital and Firstminute Capital; individual investors include Yann LeCun, Mistral's Arthur Mensch, Path Robotics' Andrew Lonsberry and Hugging Face's Thomas Wolf. The Zurich company has just emerged from stealth and builds manipulation intelligence based on force and touch, saying it can sense whole-body forces using only a robot's existing actuators, with no added sensors. The team comes from 1X, Google DeepMind, Physical Intelligence, Amazon Robotics, ETH and others. Source: AI Insider source

Maxinsights ｜ Seed ｜ $30 million ｜ Valuation undisclosedadjacent ⚠️ Vendor claim

More than 5,000 contract workers in six countries wear cameras on their heads, wrists and chests, filming first-person video while loading dishwashers, making milkshakes and washing cars; Maxinsights then processes the footage into training data for robot foundation models. The company says it has collected more than 2 million hours, of which 1.5 million hours are directly usable for training, with more than 30 customers. Founder and CEO Yuguan Yuan says all 200,000 hours of first-person data used by Genesis AI's GENE-26.5 came from it, and that it is also the largest data supplier for Dyna-2. He estimates that reaching near-human adaptability may require more than 10 million hours of diverse data. Source: Forbes source

Songyan Power (Noetix; Chinese humanoid robot maker) ｜ Hong Kong IPO ｜ Confidential filing ｜ Seeking to raise several hundred million dollarshumanoid ⚠️ Unconfirmed report

Bloomberg, citing people familiar with the matter, reports that this CATL-backed Beijing humanoid robot company has confidentially filed with the Hong Kong stock exchange (HKEX) for a listing as early as next year, with the size and timing still undecided. Founded in September 2023, the company completed a RMB 1 billion round in March this year, led by Chendao Capital, CATL's investment arm. Its products include the consumer-grade small humanoid Bumi and the N and E series biped humanoids, and it appeared on CCTV's Spring Festival Gala this year. Source: The Standard source; Bloomberg source

FORT Robotics (US) ｜ Pre-SPAC follow-on ｜ About $20 million ｜ Valuation undisclosedadjacent

On October 1, FORT announced a SPAC merger with Newbury Street II, and existing investors have now added about $20 million in further commitments. Of this, Tiger Global led about $10 million with Mark Cuban Companies participating, expected to close this week; Tiger Global separately committed $10 million to the PIPE accompanying the merger. Total committed funding tied to this deal exceeds $50 million. FORT makes safety communications and emergency stops for robots, and says its products are installed on 19,500 devices with more than 600 customers. Source: AI Insider source; Resilience Media source

Yitai Technology (Chinese spatial-data startup) ｜ Seed + angel ｜ More than RMB 100 million in total ｜ Valuation undisclosedadjacent

AgiBot and Huawei's Hubble (investment arm) were previously disclosed as shareholders; this announcement gives the combined amount and full list for the two rounds. Industry investors also include Jialichuang (JLCPCB), X Square Robot (Chinese embodied-AI startup), Zhongke Diwuji, and iSoftStone (Chinese IT services firm), and financial investors include Hillhouse (major Chinese investment firm), Shenzhen High-Tech Investment and Shanghai Semiconductor. Founded in December 2025, the company is led by founder and CEO Yang Yeqing, formerly Chief of Staff at HTC VIVE. It builds a "physical space engine" that organizes spatial structure, object states and human-hand interaction in real environments into editable, reusable spatial assets for training embodied models and world models. Source: Embodied Intelligence Home source

Tesla's Optimus Enters Production in Fremont; External Deliveries Not Until End of 2027humanoid

Tesla's Q3 update says the Fremont line has built its first "production-quality" Optimus units and the line is in the early ramp stage. The Optimus row in the capacity table has no figure filled in, and a second line in Texas is still under construction. In September The Information reported weekly output in the "hundreds"; this time the company again disclosed no output, orders or revenue. On the earnings call Musk said most of the line designed for 1 million units per year is already in place at Fremont, but the design is still changing, and 20,000 units per week is at least a year away. Remaining AI and hardware problems include how long a robot can work before wearing out, which he expects to resolve around the end of 2027, when deliveries to outside companies begin, starting with dangerous, repetitive industrial work that requires round-the-clock operation. The hand remains the hardest piece of hardware. He said "there are about 10,000 things that must be done right" to mass-produce a robot; a video of Optimus performing non-preprogrammed tasks may be released within a month or two; around April 2027 Optimus will switch to the AI5 chip, with compute about ten times AI4; and 10,000 to 20,000 robots will be needed to practice repeatedly and accumulate data in the "Optimus Academy." Berkeley professor Ken Goldberg told The Verge: "It can be built, but the software isn't enough." In the same quarter, Tesla's net profit attributable to shareholders was $787 million, down 43% year on year, while R&D spending grew about 69%. Source: AI Insider source; Huxiu source; The Verge source

Tesla: Cybercab Reaches Dallas Next Month; Owners' Personal Cars Can't Join Until Q2 2027autonomy

Tesla VP of AI Ashok Elluswamy said on the earnings call that Cybercab service goes live in Dallas next month. Owner-owned vehicles can join the Robotaxi fleet from about Q2 2027, and unsupervised FSD will then be valid only within the geofences the fleet has been approved for. Musk promised back in 2019 that personal cars could earn money by taking rides on their own, and this time the scope offered is much narrower. Installed capacity at the Texas Cybercab line is listed at more than 125,000 vehicles per year; the company notes this does not equal actual output, and all vehicles currently coming off the line go to its own fleet. There are three mileage figures: the quarterly report says unsupervised mileage reached nearly 2 million miles by quarter-end; the first Robotaxi safety report, released October 22, says cumulative mileage as of October 18 was 4.99 million km with no injuries; and Kuai Keji, citing the same safety website, gives 3.3 million miles, about 5.31 million km. The three figures use different definitions and cannot be reconciled with each other. Source: AI Insider source; Kuai Keji source; Ifeng source

Deepal SL03 Receives First Batch of Dedicated L3 License Platesautonomy

On October 22, the Deepal SL03 was issued the first batch of dedicated plates for L3 autonomous driving. The feature is only for single-lane driving in congestion on highways and urban expressways, at a maximum speed of 50km/h. At present it can only be activated on three sections in Chongqing: the Inner Ring Expressway; the New Inner Ring Expressway from Gaotanyan Interchange to Laijiaqiao Interchange; and Yudu Avenue from Renhe Interchange to Airport Interchange. On the same day, Guo Shougang, head of the Equipment Industry Department I at China's Ministry of Industry and Information Technology (MIIT), said that L3 products from Changan and BAIC received conditional access permits in December 2025 and a second batch of access pilots is under way. The current SL03 is priced at RMB 129,900 to RMB 146,900. Source: Autohome source; Sina Auto source

Maersk's Automated Fulfillment Center in Singapore Opens, Covering 1.1 Million Square Feetindustrial

World Gateway II is equipped with multi-level shuttles, an automated storage and retrieval system, autonomous case-handling robots (ACR) and AMRs, handling B2B and B2C e-commerce fulfillment across Asia-Pacific. Maersk invested more than S$200 million; occupancy is already about 70%, and the site is expected to provide about 500 jobs at full operation. Chang Kee Seng, head of contract logistics for Maersk Asia-Pacific, said the warehouse "doubles our warehousing space in Singapore." Source: Singapore Economic Development Board source

Fortescue Uses a KUKA Welding Robot to Plug and Unplug a 6 MW Charging Cableindustrial

Fortescue's 240-tonne Liebherr T264E battery-electric haul truck has a 3.2 MWh battery, roughly equal to 40 Model Ys. The company's goal is a full charge in 30 minutes, and the 6 MW liquid-cooled cable it uses is thick and heavy, impossible for a person to handle. Fortescue adapted a KUKA welding robot, normally used in car factories, to plug and unplug the cable; the 6 MW fast-charging test rig received A$10 million in funding from the Australian Renewable Energy Agency in 2024. Fortescue says in the video: "Nobody makes a 6 MW charger, so nobody makes a lot of the components either." Source: Electrek source

MBody AI Signs Caribe Royale Resort in Orlandoadjacent

This is Nasdaq-listed MBody AI's first customer in Florida and its first resort customer not tied to a casino, billed as a multi-year subscription. Its Orchestrator software coordinates autonomous devices from different vendors. Source: Taiwan News source

Agility Names Nvidia the Sole Onboard Compute Supplier for Digit 5humanoid

On October 21, Nvidia announced it was bringing its Halos safety system to Agility. The next day Agility announced that Digit 5's onboard compute will standardize on the NVIDIA IGX Thor module, making Nvidia the sole onboard compute supplier for the production version of Digit 5. Digit 5 will also enter Nvidia's Halos AI Systems Inspection Lab, where safety software, AI components and cybersecurity are evaluated ahead of final third-party certification. Early supply is set for H1 2027, followed by a gradual ramp over 6 to 9 months, in line with the 2027 pilot at AGCO's Kansas plant. CTO Pras Velagapudi said: "Moving to the Nvidia Thor module gives us a common foundation to raise Digit's capabilities and prepare for large-scale commercial deployment." Agility's merger with Churchill Capital Corp XI is expected to close in Q4. Source: GlobeNewswire source

Unitree: About 10,000 to 20,000 G1 Units Built So Far, and Tens of Millions of RMB a Year for a Physical AI Competitionhumanoid

Unitree chairman Wang Xingxing said at the 2026 World Physical AI Conference in Hangzhou that the G1 has been produced and shipped off the line in a cumulative 10,000 to 20,000 units. Unitree said in early June that as of May a single biped humanoid model had cumulatively come off the line at about 11,000 units; this time only a range was given, and its lower bound is below the figure from five months ago. He also announced that the World Physical AI Competition will be funded with prize money of tens of millions of RMB per year. ⚠️ Speech claimSource: Sina Finance source

A Factory Depalletizing Robot Project Quoted at Over RMB 1 Million for Hardware and Softwareadjacent

During this year's China International Industry Fair, an embodied-AI secondary-development service provider worked through the numbers for the column "Radio Wave No. 42": a depalletizing project with a single robot unit plus development and delivery is quoted at more than RMB 1 million for hardware and software combined. The company has fewer than 20 people, and the project value it has taken on this year has reached the tens of millions of RMB, involving dozens of robots. Robot-dog projects have thin margins: an order totaling about RMB 200,000 may earn only RMB 10,000 to RMB 40,000. Robot makers often subcontract development; a salesperson at a secondary-development firm said that in some partnerships the robot maker "wants to earn money on your hardware and also take 10% of your software." AgiBot founder, chairman and CEO Deng Taihua disclosed in April that the partner channel accounted for 25% of sales in 2025, with a target of 60% this year. Source: Huxiu source

AgiBot Leases Over 220,000 Square Meters of Office Space in Shanghaihumanoid

AgiBot and affiliated companies signed three projects under Lujiazui Group, Qiantan Yuan, Qiantan Shang and Yuqiao Smart Valley, with a combined area of more than 220,000 square meters. A person close to AgiBot confirmed to STAR Market Daily that the news is true. Shanghai's previous single office-lease record was about 90,000 square meters, leased by Hypergryph (Chinese game developer) this August. At market rates, the annual rent is estimated at about RMB 350 million, excluding property management fees. Source: Sina Technology source

XPeng's IRON Re-infers Every 0.2 Secondshumanoid

XPeng has deployed the temporal prediction capability of its second-generation VLA on IRON's onboard compute: every 0.2 seconds it re-infers from what is in front of it, while predicting action changes over the next 6 seconds. In the demo, when a person suddenly switched from an open palm to a fist, or changed where to high-five, IRON changed its motion accordingly. XPeng's plan remains mass production by the end of 2026 and deliveries starting in H1 2027, going first into stores and shopping malls. Source: Sina Technology source

China Telecom Incubates Humanoid Robot Company Gamma Intelligencehumanoid

Gamma Intelligence, the first market-oriented company incubated by the China Telecom Institute of Artificial Intelligence (TeleAI), has registered in Xuhui, Shanghai, and is based at Mosu Space. Its research focus is "self-evolving" humanoid robots and whole-body models, meaning locomotion, balance and dexterous manipulation in a single unified model. The investor is the China Telecom (Shanghai) AI Industry private fund, with the amount undisclosed. Bai Chenjia is founder and CEO, and China Telecom CTO Li Xuelong serves as scientific advisor. Source: Xinmin Evening News source; Pandaily source

Chery's Mojia Launches Consumer Humanoid "Mobao"humanoid

"Mobao" stands 0.9 meters tall and has 22 degrees of freedom. It was launched on October 21 with global pre-orders open, and is expected to enter mass production and delivery by year-end, in high, mid and low configurations. According to Beijing Youth Daily, the price may be tens of thousands of RMB. Zhang Guibing, Chery executive vice president and general manager of Mojia Robotics, said the aim is to "use the cost logic of building cars to make robots that can be mass-produced, delivered and profitable," and said Mojia Robotics' global sales have exceeded 3,000 units, reaching more than 60 countries and regions. ⚠️ Vendor claimSource: Beijing Youth Daily source

Schaeffler and Comau Bring Nvidia Simulation into the Factoryindustrial

Schaeffler has signed a technology partnership with Germany's Wandelbots, planning to make the Wandelbots NOVA platform the robot development standard across Schaeffler's global plants. New applications are first tested virtually in Nvidia Omniverse and Isaac Sim, then replicated to other plants once validated. The two companies say the full lifecycle cost of automation solutions can fall by up to 40%. Wandelbots CEO Christian Piechnick said the past 12 months have proven the technology is feasible, and the task now is to make skill development efficient, transferable across systems and reliable on the production line. At GTC Berlin, Italy's Comau announced it is putting Isaac Sim digital twins, the SimReady asset library and Jetson compute into its next-generation robot motion controller. ⚠️ Vendor claimSource: MarketScreener source; Tech Times source

China's L3: Progress on Access, Marker Lights and License Plates Within One Weekautonomy

On October 20, MIIT sought public comment on the "Installation Requirements for Exterior Lighting and Light-Signaling Devices on Automobiles and Trailers," which stipulates that ADS marker lamps may only be fitted to L3-and-above vehicles, so the "little blue light" common on L2 models will be banned. On October 22, Guo Shougang, head of MIIT's Equipment Industry Department I, said at the World Intelligent Connected Vehicles Conference that a second batch of L3 access pilots is under way. He judged that autonomous driving is "in a critical period of moving from usable to user-friendly," and said vehicle operation monitoring will continue to be strengthened. On the same day, the Deepal SL03 received the first batch of dedicated plates (see the commercialization section). The mandatory national standard behind these moves, GB 44721—2026, was approved and released on July 30 and takes effect on July 1, 2027, requiring L3 systems to monitor the driver's ability to take over. Source: China News Service source; Sina Auto source; Cailian Press source

Hubei and Singapore Use Challenge Problems to Find Real-World Scenariosembodied

On October 22, the Hubei Humanoid Robot Innovation Center released more than 20 embodied-intelligence "open bidding" (揭榜挂帅) projects across six categories, with a reward of up to RMB 500,000 per project. The problem statements are very specific: the robot must still stably complete steel-bridge welding when the workpiece position is offset up and down by 15 centimeters, for no fewer than 4 consecutive rounds; elder-care tasks require understanding dialects and recognizing elderly people's emotions; and another has a robot make Hubei-style hot dry noodles and sanxian doupi end to end, from ingredient prep to serving. Liu Chuanhou, chief operating officer of the innovation center, said "all the problems on the list come from real needs in society." The list is tied to the on-site practical training special action launched in June by MIIT and the State Council's State-owned Assets Supervision and Administration Commission, which aims to form more than 100 high-value scenarios by year-end. The same day, Singapore's A*STAR and Micron launched the first national embodied-AI manufacturing challenge, with total prize money of S$88,000; the task is to carry out disassembly and assembly, inspection and cleaning in a semiconductor chamber mockup built by Micron, with the finals on March 20, 2027. Source: Jimu News source; TNGlobal source

Hands and Actuators: Orders and Capacity Begin to Divergehardware

On Tesla's Q3 call, management said the design of most Optimus electromechanical systems has converged, but the hand remains the hardest hardware, with the difficulty in custom gears, structural design and long-term reliability. Huxiu relays Musk's remark that final output depends on the weakest link in the supply chain. The component makers' books show the other side. A brokerage research report dated October 16 and reposted by Gelonghui summarizes three types of humanoid components at Zhongding Group (Chinese auto-parts maker): harmonic reducers with annual capacity of 150,000 units and cumulative orders of about 300,000; force sensors with annual capacity of 200,000; and micro motors with annual capacity of 540,000 and cumulative orders of about 180,000. The designated customers are EngineAI (Chinese humanoid maker), Fourier (Chinese humanoid maker) and an unnamed leading company. Cumulative harmonic reducer orders equal two years of capacity, while micro-motor orders are only one third of annual capacity. For dexterous-hand component maker Zhaowei, embodied-robot revenue of RMB 23.87 million in 2025 is still under 2% of total revenue. Source: Huxiu source; The Verge source; Gelonghui; Sohu source

Robot Onboard Compute Is Concentrating on a Few Platformshardware

This week Agility named Nvidia the sole onboard compute supplier for the production version of Digit 5. Tesla plans to switch Optimus to AI5 around April 2027, which Musk says has about ten times the compute of AI4. LG Innotek built its humanoid vision solution on Nvidia's Holoscan Sensor Bridge. In China, Huixi Intelligence has raised more than RMB 2.5 billion this year, and its Guangzhi R1 platform is already used on the AgiBot Genie G2. Robot makers and sensor suppliers outside China are setting their interfaces around Nvidia's platform; Chinese robot makers have one more option, a Chinese chip, though Huixi's publicly disclosed mass-production case is currently only AgiBot. Source: GlobeNewswire source; AI Insider source; Korea JoongAng Daily source; 36Kr source