Today's Highlights
· Perceptron open-sources 36-billion-parameter embodied foundation model Isaac 0.5, single-demonstration learning outpaces π0.5
· Meta lets robots into AI data centers to swap cables, one employee estimates it could handle 80% of the work
· Caocao Mobility (Chinese ride-hailing platform) signs Abu Dhabi's K2, set to deploy its first overseas Robotaxi fleet this year
· NSF commits $30 million for a 39-person, six-university study on how humans and robots adapt to each other
· Xuanchuang Robotics (Chinese industrial robot startup) wins an order of over 100 units from PetroChina, delivering 20 per month, alongside a closed Series A1
· Robot at Xiangyang Science and Technology Museum injures girl by kicking her, operator reaches settlement with the family
Research Progress
Let the VLM point at a spot on the image, leave the geometry to SLAM · autonomy
At each step, the robot receives four surround-view images; rather than having the vision-language model compute 3D coordinates, it only needs to pick one of the images and mark a single 2D pixel point. The depth map and camera intrinsics then convert that point into a 3D position, which is handed off to the underlying SLAM system. Using this approach, the success rate on R2R-CE was pushed up to 66.2%. A Zhejiang University team argues the problem is that VLMs, from pretraining onward, only ever see 2D images and text — forcing them to compute 3D coordinates instead produces spatial hallucinations. In an ablation experiment they reverted to conventional metric waypoint output, and success rates on R2R-CE and RxR-CE dropped by 24.8 and 25.2 percentage points respectively, confirming this hypothesis. The approach also saves considerable overhead: each trajectory needs an average of only 9 interaction steps, taking 16.58 seconds per task, compared with 37.47 seconds for StreamVLN and 41.46 seconds for DualVLN. Chain-of-thought reasoning isn't run continuously either — it only triggers at key points like intersections or doorways, accounting for about 30% of the trajectory.
Hongyan Feng et al. (Zhejiang University ACES Lab × Zhejiang Humanoid Robot Innovation Center) · arXiv 2608.17512 source · Analysis: TechWalker source
Robotic fish builds a light field while questioning its own location · perception
Once underwater gliders submerge, they lose GPS signal and must rely on dead reckoning to estimate their position — a method whose error grows the longer it runs. A team at Michigan State University's approach is to treat this uncertainty itself as useful information: a multi-fidelity Gaussian process scores each measurement, and the less confident the robot is about its current position, the lower the weight assigned to that data point. They applied this approach to reconstructing a 3D light field in a large indoor water tank, bringing the final mapping error down to less than a quarter of that from conventional methods.
Demetris Coleman et al. (Michigan State University · JHU APL) · Autonomous Robots · Analysis: Bioengineer.org source
Prosthetic e-skin measures pressure and temperature simultaneously, modules snap together like Lego · perception
The fabrication process starts by scanning the prosthetic's shape, then laying out the sensor array to match that geometry, and finally using 3D printing combined with laser cutting to produce thin-film modules. The modules snap together via interlocking joints, with no adhesive needed. A Washington State University team says the system's sensing resolution reaches 10 times that of existing commercial glove sensors, and maintains high coverage density even on curved freeform surfaces. First author Hongyi Shen said this advance "makes the production of medical-grade e-skin accessible," laying the groundwork for a fully bionic skin that can both sense and deliver tactile stimulation.
Hongyi Shen et al. (Washington State University) · Cell Reports Physical Science · Analysis: Futurity source
Other papers today: ITRM uses trajectories "imagined" by a world model to give reinforcement learning agents early warning of environmental anomalies, without needing to inspect the policy internals — achieving an average AUROC of 0.853 across navigation, robotic control, manipulation, and autonomous driving tasks, with a false-positive rate of only 0.279 at a 95% true-positive rate.
Open Source · Tools · Benchmarks
· Isaac 0.5: Perceptron has open-sourced a 36-billion-parameter embodied foundation model, with weights on Hugging Face and fine-tuning code and technical documentation on GitHub. Training used over 3 trillion multimodal tokens, including 1 million hours of general video and 100,000 hours of robot experience, covering more than 35 robot embodiments. The company says the model integrates video understanding, embodied reasoning, and action generation into a single system; it scores 97.2% on the LIBERO benchmark, roughly on par with GR00T N1.7's 97.0% and π0.5's 96.9%. The real gap shows up in learning new tasks: given just one demonstration, error improves by 7.0–10.5x, versus only 2.3–3.1x for π0.5. The company was founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both formerly of FAIR. source
· Benyuan Zhishu's "Dual Billion Plan": At the Big Data Expo, the company announced a five-year plan to build 1 billion expert trajectory data points plus 1 billion hours of embodied manipulation data, while also releasing four platforms: OriginBench for evaluation, OriginWise for expert data production, OriginFlow for agent operation, and OriginPhysics for embodied data collection. Founder Li Niyong said customers are no longer asking "how much data do you have" but rather "what kind of data will keep the model improving." Note that these figures are the company's stated targets, not its current scale. source
Funding and Deals
Xuanchuang Robotics ｜ Series A1 ｜ Tens of Millions of Yuan · industrial ⚠️ Company account
The investors are Qianhai Fangzhou, Guangyang Co., and West Lake Venture Capital; the funds will go toward data pipelines, training and deployment of VLA and JEPA world models, and stocking standardized products. Founded in December 2022, the company's core team comes from Harbin Institute of Technology and specializes in explosion-proof inspection for hazardous oil and gas/chemical settings, following an "inspection plus operation" approach. Its configurations include wheeled, rail-hung, tracked, and dual wheel-legged designs — notably, no humanoid form. Founder and CEO Fu Zhe disclosed that PetroChina's procurement approach has shifted from POC-project purchasing last year to bulk purchasing this year; an order of over 100 units signed with PetroChina's Xinjiang Oilfield division has already completed its first batch delivery, with subsequent deliveries proceeding at a pace of 20 units per month, and current orders on hand already exceed 100 million yuan. The first standardized final-assembly line came online on July 30, built jointly with Guangyang Co. Asked why the company doesn't build humanoid robots, Fu explained that chemical plant customers prioritize reliability far above form factor, and that a wheeled chassis with dual arms is "more reliable, more economical, and easier to deploy."Source: 36Kr source
Green Harmonic Drive (688017.SH) ｜ Plans H-Share Listing in Hong Kong ｜ Amount Undisclosed · hardware
The company announced on the evening of August 26 that its board has approved a proposal to issue H shares and list on the main board of the Hong Kong Stock Exchange, pending approval from shareholders, the China Securities Regulatory Commission, and the Hong Kong Stock Exchange. First-half revenue reached 349 million yuan, up 38.64% year-on-year; net profit attributable to shareholders was 70.11 million yuan, up 31.25%; non-GAAP net profit was 58.17 million yuan, up 36.92%. The company attributed this growth to an increased share of the industrial robot market and significant expansion of its embodied AI robotics business. Harmonic reducers and metal components contributed 288 million yuan, or 82.67% of revenue, with a gross margin of 32.79%; electromechanical products contributed 54.74 million yuan. The company was the earliest in China to achieve industrial-scale mass production of harmonic reducers, and holds roughly 15% global market share, ranking second worldwide and first in China.Source: Gasgoo source
EXL (EXLS.O) Acquires iMerit ｜ $310 Million ｜ Closed August 3 · adjacent
The deal consideration is split into two parts: $170 million upfront, plus up to $140 million in milestone-based earnout payments over the next two years. iMerit provides data services for AI model training, evaluation, and reinforcement learning, covering text, image, video, audio, and LiDAR data, serving robotics, autonomous driving, and other autonomous systems customers. Founder and CEO Radha Ramaswami Basu will become Executive Vice President of EXL and join its executive committee.Source: The Robot Report source
Commercialization and Deployment
Kinisi's KR1 installs crankshaft bearing shells at a UK auto plant, trained on just one day of human demonstration · embodied ⚠️ Company account
The bearing shells must be pressed into the crankshaft housing bore, with the press-fit force varying between 2,000 and 8,000 newtons depending on engine specification; if the fit is off, a camera alone can't detect it — only a change in resistance reveals the problem. This wheeled humanoid robot now handles three operations: picking and placing crankshaft components, fitting the bearing shells into their corresponding grooves, and completing precision-aligned press-fitting. It stands 162 cm tall and weighs about 100 kg; the customer is an unnamed global automaker, with the pilot running at Kinisi's headquarters in Bristol. CEO Brennand Pierce told UK tech outlet IZON that the automaker has identified over 15,000 similarly automatable operations across its global plants, with this single customer alone representing an addressable opportunity of over $2 billion; South Korean industry outlet TheGuru cited this report on August 27. Kinisi was acquired by BearRobotics (LG-owned robotics company) just ten weeks ago. Most previously publicized humanoid robot demonstrations have focused on warehouse logistics tasks, whereas precision engine component assembly has long served as a litmus test that automotive engineers use to judge whether embodied AI truly delivers manufacturing value.Source: Tech Times source
Caocao Mobility signs Abu Dhabi's K2, set to deploy first overseas Robotaxi fleet this year · autonomy ⚠️ Planned, unconfirmed
K2 is an AI technology company owned by the Abu Dhabi government; under this partnership, K2 will provide the testing environment, operational infrastructure, and local resources, while Caocao Mobility will supply intelligent customized vehicles, technical support, and training. The two sides plan to begin local testing and validation in 2026. This builds on Caocao's earlier cooperation with the Abu Dhabi Investment Office. Back in China, driverless testing without a safety operator began in Hangzhou's Binjiang district on July 27, using a station-triggered dispatch model with a local fleet of about 100 vehicles in Hangzhou, backed by Caocao Mobility's RAS remote safety service platform, which handles continuous monitoring, emergency call response, and remote takeover; back in April, it had already become the first company approved for driverless road testing in Hangzhou. China's first ground-up-designed Robotaxi model, the Eva Cab, has also been unveiled, with mass production set for 2027. Under the goals laid out in its RoboX strategy, the company aims to deploy a cumulative 100,000 Robotaxis and 100,000 Robovans by 2030.Source: scanx.trade sourcesource
Einride to put 500 Tesla Semis on the road, rolling out across five states starting in September · autonomy
Texas, California, New Jersey, Illinois, and Georgia will be the first five states to gain access, with the trucks running on Einride's own fleet intelligence platform; customers include shippers at the scale of Amazon. The full rollout will proceed in phases over the next two years. Tesla Semi Program Director Dan Priestley said electric heavy-duty trucks lower per-mile costs through better fuel efficiency, reduced maintenance, and higher uptime, and that the economics will improve further once deployment scales up. Tesla has set the Semi's production launch event for September 24 at its Nevada factory, though large-scale deliveries of the vehicle have already been promised for several years.Source: CDLLife source
Tesla Robotaxi launches in Miami with no safety monitor on board · autonomy
Service currently covers only a small area of West Miami, with the map Tesla posted on X quite tightly bounded. This mirrors how the service launched in Houston and Dallas in April, both starting from very limited coverage areas, while Austin only expanded to citywide coverage last month.Source: Mashable source
Industry Developments
Meta puts robots to work in AI data centers, testing cable swaps and server reboots · industrial ⚠️ Single-source account
The robots being tested come from three manufacturers: San Francisco-based Watney Robotics, Quebec-based Kinova, and Zurich-based ABB, handling tasks including moving racks, inspecting equipment, swapping network cables, and rebooting servers. According to WIRED, an unnamed employee estimated that a robot capable of reliably completing cable-swap tasks could replace up to 80% of the work for certain roles — though this is that employee's personal estimate, not an official Meta forecast, and the robots currently still work more slowly than humans. Obstacles, battery life, visual inspection, and dense cabling remain sticking points, and the process still requires human oversight throughout. In response, a Meta spokesperson said the US is in the midst of its largest infrastructure boom since World War II, facing a severe shortage of skilled workers, adding: "We need more workers, not fewer."Source: Decrypt, citing WIRED source
NSF commits $30 million to establish center on human-robot co-adaptation, led by UT Austin · embodied
The five-year project brings together 39 researchers across six universities, with a research focus on long-term coexistence: how both people and robots change after living alongside one another for months or even years. Center director Joydeep Biswas, associate professor of computer science at UT Austin, said the next real leap "isn't just about getting robots to perform more tasks, but about getting them to understand the people around them — their needs, preferences, values, and constraints." The research relies on a real-world deployment network called HERO, with sites set up in homes, dormitories, cafes, public museums, rehabilitation hospitals, and senior/assisted-living residences at UT Austin, MIT, Yale, Indiana University Bloomington, the University of Utah, and Tufts University, using long-term deployment to observe what happens once the novelty wears off. Industry partners involved include Amazon, Apptronik, Diligent Robotics, Google DeepMind, Hello Robot, MassRobotics, NVIDIA, and Robust AI.Source: Weidu.net source
Over 70 embodied AI training grounds completed and operating nationwide in China, with 46 more under construction or planned · adjacent
This figure is as of the end of June 2026. Guizhou's first embodied AI "data factory" recently began operations, featuring 16 fully realistic simulation scenarios across 6 major categories — research, dining, industrial, retail, home, and office settings — following a complete pipeline of "scene capture → equipment calibration → cleaning and annotation." Zhang Jianwei, a foreign academician of the Chinese Academy of Engineering, said at the Big Data Expo that large-scale deployment of embodied AI is at a critical "window period," requiring data companies and robotics teams to jointly conduct training and validation in order to bring down innovation costs. At the same exhibition, China Southern Power Grid showcased a robotic dog rated to carry 200 kg of payload and climb 60-degree slopes.Source: China News Service source
Robot at Xiangyang Science and Technology Museum injures girl by kicking her, operator reaches settlement with family · humanoid
The robot was in a calibration/rehearsal phase at the time of the incident. Hefei Anda Chuangzhan Technology Co., the museum's operations service provider, said the compensation details would not be disclosed publicly; the robot involved was a Unitree unit. According to the museum, Unitree robot performances are still being held regularly.Source: Southern Metropolis Daily source
JD.com acquires land in three cities within two months to build RoboBases, targeting 80 within five years · adjacent ⚠️ Planned, unconfirmed
On August 26, JD.com won a bid for an industrial plot in Wuxi's High-Tech Zone, with a planned floor area of about 112,000 square meters; earlier, a Guangzhou Huangpu project broke ground on July 11 with a planned area of 190,000 square meters, expected to be completed by the end of 2028; a plot in Shanghai's Pudong Zhangjiang area spans 105,000 square meters. Each RoboBase is meant to combine display and delivery, maintenance, R&D and design, pilot-line assembly, data collection, and manufacturing functions all in one; iiMedia Research founder Zhang Yi likened it to a robotics "4S dealership," and said outright that JD.com's bet isn't primarily on hardware R&D itself. Zheng Xiaodan, head of JD Retail's embodied AI business, said the robotics industry boom isn't a moment you can wait for — it has to be actively built.Source: National Business Daily source
Galbot (Chinese embodied AI startup) unveils its first bipedal humanoid ET1, targeting a price of about $95,000 · humanoid ⚠️ Company account
After three years of building wheeled robots that pick items in pharmacies and convenience stores for handoff to delivery riders, ET1 marks the first time this Beijing-based company has put a robot on two legs. The unit stands 173 cm tall and weighs about 65 kg, with 48 degrees of freedom overall, including 24 across both hands; its compute platform is NVIDIA's Jetson Thor, and it carries an IP54 protection rating. It has already demonstrated fully autonomous play against human athletes in a sporting event (previously reported); the company's new figures put reaction time at 0.1 seconds, achieving roughly a 90% forehand contact rate against incoming balls traveling over 50 km/h. TechEBlog notes that most of these specifications come from aggregated data sheets rather than an official full spec release, and that the company itself still regards ET1 as a prototype.Source: TechEBlog source
ICE plans to spend up to $2 million on robot dogs · adjacent
The stated purpose is use in immigration enforcement operations. The figure comes from a government funding document reviewed by NBC News.Source: NBC News source
Hardware · Supply Chain
· Harmonic reducers: industry analysis estimates these account for about 30%–35% of a humanoid robot's total cost, concentrated mainly in lighter-load joints such as the wrist, elbow, and fingers source
· Jinlu Electronics (301282): PCBs used in integrated joint modules for humanoid robots have moved from trial production to mass production; the company stated on an investor platform that this business still accounts for a small share of revenue and has not yet had a material impact on results source
· Laifu Harmonic Drive: robot component revenue grew sharply, and adjusted losses narrowed source
Top comments (0)