Persona AI explains that its humanoid robots target skilled trades in heavy industry rather than warehouse handling: the shipbuilding sector faces a labor gap of about 40,000 workers, welding shortages constrain production, and shipyards are designed for human bodies, so legged robots require no site modifications.
Shi Ma Notes interviewed Hu Yucheng and Guo Yanjiang, the two lead authors of VPP (PhD students of Professor Chen Jianyu at Tsinghua University's Institute for Interdisciplinary Information Sciences), covering three works: VPP, BagelVLA, and Ctrl-World.
Are we living in a simulation? If you are a robot, the answer may be yes.
#India once fought to shape its own destiny; today, we are building it through engineering. 🇮🇳🤖 At @kody_technolab, independence means building homegrown #robotics and #AI to help a self-reliant, technology-driven nation. Happy Independence Day, India! ✨ #IndependenceDay2026 #Roboticscompany #KodyTechnolab
Physical Intelligence co-founder Chelsea Finn spoke at Startup School 2026 about building general-purpose robots in the real world: reinforcement learning doubled robot throughput, and the system can run autonomously for hours. She believes robots are entering the "GPT era"—moving from specialized models to general systems that work across tasks and across robots. Sergey Levine remarked that Chelsea's talk was once again rock-star level, and that robots really can fold laundry and make espresso!
Skydio compares its Drone as First Responder (DFR) with Flock's automated license plate recognition (ALPR) network point by point: Flock cameras run continuously, record every vehicle that passes, and retain data for 7 days by default, with city and county deployments ranging from dozens to thousands of units. DFR, by contrast, collects data only when a named human pilot launches for an authorized police, fire, or EMS mission, and video is uploaded to a secure digital evidence management system; most cities and counties currently have fewer than 10 DFR drones.
Insta360 CEO Liu Jingkang said the company has maintained a panoramic camera market share above 60% for years, but it remains two orders of magnitude away from phones, and the company strongly opposes vicious price wars. Insta360 launched its new ultra-flagship panoramic camera X6 on August 12, and also previewed the Go Ultra with AI voice features. Currently 43% of Insta360's code is AI-generated, and the company's goal is to improve efficiency rather than pursue being AI native.
In an interview, Li Zexiang described his incubation system, from the Hong Kong University of Science and Technology's 3126 Laboratory to the XbotPark robotics base in Dongguan's Songshan Lake in 2014. As of 2025, it has incubated more than 280 hard-tech companies with a combined valuation exceeding 500 billion yuan, of which 4 have gone public and 12 have grown into unicorns or quasi-unicorns.
On August 8, the production team of "Thoughts Shine on Lingnan" visited the XbotPark robotics base in Songshan Lake, Dongguan, and spoke with founder Li Zexiang to analyze the Greater Bay Area's supply chain and science and innovation ecosystem. Li Zexiang said that hardware iteration in the Greater Bay Area can be 20 to 30 times faster than in Silicon Valley, at only one-tenth the cost, and more than 40% faster than in the Yangtze River Delta; in XbotPark 1.0 (2014 to 2020), 60 small teams had an 80% survival rate, and 15% became unicorns or quasi-unicorns.
At its Advancing AI event, AMD introduced the Ryzen AI Embedded X100 series processors for humanoid robots, the Kria AI SOM modules, and the Robotics Software Suite, with sampling in June and mass production in Q4. The article's analysis holds that AMD's combination of x86 real-time control, unified memory, and FPGA targets industrial quasi-humanoid and dual-track supply chain scenarios; in the short term it is unlikely to shake Nvidia Thor's roughly 68% share of the high-end bipedal segment, and its opportunity lies in capturing the installed base of industrial PCs plus discrete GPUs and serving as a second supplier.
Chen Tao, director of Fudan University's Deep Learning Laboratory, said in an interview with The Paper's tech section that mainstream VLA models rely too heavily on successful-trajectory data and lack failure and error-correction data, leaving robots without autonomous error correction or continual learning capabilities. The embodied foundation model architecture urgently needs a breakthrough, and an embodied GPT moment will be hard to achieve within two years.
In an interview, Boston Dynamics mechanical engineer Taylor Frey-Baker discusses the design of the electric Atlas's head and how she entered the field. She is responsible for the compute core inside the head, which houses four perception modules with stereo vision; the head's modular design allows the entire unit to be swapped out and the compute core upgraded like a game cartridge. The hydraulic Atlas did not need a head because it performed movement demonstrations, whereas the electric Atlas relies on head cameras for a wider field of view and uses lights and sound to signal its status to people nearby.
Founded in 2004 and active in collaborative robotics for more than 20 years, PAL Robotics' mobile manipulation platform TIAGo Pro is cited as a success case for Catalonia in ACCIÓ's report "La robòtica a Catalunya"; the platform uses flexible-drive robotic arms and an enhanced non-verbal interaction system.
Thanks for having us at the Advancing AI event last week, @AMD! Next year will be a big year 🤫
Ding Wenchao, co-founder and chief scientist of Tashi Intelligence, judges that embodied AI is currently close to the level of autonomous driving in 2021-2022, and may enter the 3.0 era with in-context learning capability in about 24 months. At WAIC, Tashi Intelligence released its new-generation embodied AI native foundation model AWE 3.5, which greatly increases data volume compared with 3.0, moves toward a native model architecture, and collects data in a human-centric way with tactile gloves, working with a 21-DOF dexterous hand to complete manipulation.
AI is not the hard part. Getting it to run consistently and reliably across dozens of factories worldwide? That is the challenge. Magna's Todd Deaville joined the @RoadToAutonomy show to talk about AI, robotics, and what it really takes to scale innovation in manufacturing. 🎧 https://bit.ly/4pEOPNJ
At the 2026 WAIC Embodied AI Forum, six guests predicted that robots' ChatGPT moment is two to five years away, with Yao Maoqing saying two years and Xu Danfei saying five, though no consensus has formed on the criteria for judging it. The discussion focused on three main threads: the data flywheel must be built through real deployments rather than data collection alone; VLA and World Action Models (WAM) are not opposites, with PI continuing the VLA path through π0.7 while Dyna turns to WAM; full-stack capability is becoming a trend, but a full-stack business model is not necessarily the only answer.
The latest installment of BAAI's Scholars Speak takes AI self-evolution as its topic, with five scholars—Sui Yanan, Gong Jiangtao, Zhou Shiji, Yang Jian, and Li Chaozhuo—discussing it from perspectives including embodied intelligence, cognitive-augmentation agents, agent decision optimization, AI coding, and trustworthy AI safety.
Yunqi Capital released a transcript of a roundtable at Tencent WAIC Night, hosted by Managing Partner Chen Yu, in which Huang Xiaohuang of Manycore, Luo Yihang of Shengshu Technology, and Zhang Zhizheng of Galbot discussed the next paradigm after Agentic AI.
Why it matters: The three guests' disagreements over the world-model approach and their differing judgments on the pace of deployment offer a multi-party reference for understanding the Physical AI debate.
At the "Intelligent Trends Forum" at WAIC 2026, Cao Yunan, Chairman of Elite Robots, delivered a speech titled "Optically Connected Intelligence: The Strategic Outpost of Physical AI," noting that physical AI deployment faces four bottlenecks: data, models, execution, and scale. The company's self-developed "Yuanqi Primo" solution enables on-site cold start with a small number of samples, keeps sensitive data resident locally, and evolves through federated intelligence; it has already achieved scaled deployment in optical communications manufacturing. Its Suzhou super factory is expected to be completed and put into operation by the end of 2026, and the company has accumulated millions of robot operation trajectories and hundreds of millions of force-tactile sensor sampling data points.
Yunqi Capital published a talk on July 17, when WAIC 2026 opened, titled "The Year of Action," using five numbers to observe AI moving from "answering questions" to "completing tasks." As of early July, OpenClaw had received 380,000+ GitHub stars; after the same model was switched to a more complete Harness architecture, the task success rate on a coding Agent benchmark rose from 71.75% to 81.36%.
Yin Qi delivered a keynote titled "When Agents Enter the Physical World" at the WAIC 2026 opening ceremony main forum, proposing the formula "agent capability = model capability × agent operating system capability." The talk centered on three structural shifts in the agent era—new systems, new devices, and new networks—and noted that an Agentic OS and the A2A network will become key infrastructure. StepFun is conducting frontier research with the Shanghai Qi Zhi Institute, led by Turing Award winner Andrew Yao.
Ars Technica interviews MIT's Vincent Sitzmann, Runway CTO Anastasis Germanidis, and World Labs co-founder Ben Mildenhall to explain the definition of world models, the diverging technical approaches, and the questions that remain unsettled.
Every grasp is data. Every failure is feedback. Every iteration makes Daksha smarter. Humanoid intelligence is being built through action. #Daksha #HumanoidRobot #EmbodiedAI
Yunqi Podcast Attent!on speaks with Wang Ximing, co-founder of the brain-computer interface project Kongshanci, about its approach of using brain-computer interfaces and neuromodulation to address psychiatric and cognitive disorders: the company was founded in October 2024 and has completed angel, Pre-A, and A rounds of financing (with Yunqi continuing to invest); it currently has three products, one of which has entered clinical use at hospitals, one is a home-oriented sleep product combining EEG with electrical stimulation (internal testing completed), and another ultrasound brain-computer interface technology is in development.
Union Square Ventures believes that as the cost of models handling messy unstructured input falls and sensors and hardware become widespread, data loops are expanding beyond software into areas that were previously out of reach. Ambient speech, once transcribed and structured, becomes a usable dataset, giving rise to vertical applications and horizontal tools such as Abridge and Granola; in the physical world, robots are becoming more capable and cheaper, and related scenarios are moving from experiments toward commercialization.
Ars Technica interviewed several robotics researchers and founders, outlining how AI is driving the development of general-purpose autonomous robots and the data, safety, and reliability obstacles they face.
KEENON Robotics announced it will attend Interclean Amsterdam 2026, presenting cleaning service robot solutions related to sustainable intelligence. The body text on which the summary is based is actually an industry analysis: China is already the world's largest and fastest-growing robotics market, with the service robot market exceeding 28 billion RMB and growing nearly 40% year on year; funding jumped 344% year on year in 2020, and in the first half of 2021 most leading manufacturers were at B/C rounds, with funding mostly between 300 million and 1 billion RMB.
We have been approaching reward supervision for robots the wrong way. I think free-form preferences are part of the answer. Short thread 🧵
The Diffusion PR team spent four days at the Automate 2026 trade show in Chicago, where they saw robot baristas making coffee on site. Physical AI was the buzzword of this year's show, and Alexandre Boffi of Kawasaki Robotics explained it as the mechanical embodiment of artificial intelligence in the physical world.
Dexmate argues in a post that the deployment ceiling for industrial Physical AI is set by hardware rather than models, using its own Vega 1 Pro as an example.
Enchanted Tools presents the application positioning of its humanoid robot Mirokaï for real commercial environments, with target scenarios covering healthcare, senior living, retail, museums, hospitality, and public spaces. Mirokaï is designed to take on repetitive reception tasks such as greeting visitors, answering questions, and guiding people through buildings, so that staff can devote their time to interactions that require a human presence. This is promotional product content; the original text discloses no technical specifications or verified deployment results.
Nobody buys a robot. People buy back time. Caregivers' time. Employees' time. People's time. → This is the true original design intent behind Mirokaï. ⊕ More details: https://enchanted.tools/robot ⊕ Book a meeting: https://enchanted.tools/contact
StridingAI introduces its Physical AI vision in animated form, in which multimodal intelligence drives real-world robotic action. The company aims to connect processes from inventory management to customer service, improving efficiency in retail, hospitality, industrial, and everyday environments.
The chapter "Trends in Open-Source Commercialization" from the China Open-Source Development Deep Report (2025), written by the Yunqi team, has been officially released. It notes that global open-source commercialization is moving from a service-oriented model into an intelligence- and ecosystem-oriented one. In 2025, global commercial open-source software companies raised about 6 billion USD, of which open-source AI accounted for about 3.8 billion USD and China about 1.1 billion USD. The report adds dedicated observations on the commercialization of open-source AI hardware and embodied intelligence, noting that open-source robotics commercialization unfolds along three main lines: hardware sales, operations and maintenance services, and scenario-based solutions.
The article discusses how teleoperation technology could move work once considered necessarily local, such as excavator and loader operation, into the global labor market.
Our message at @VivaTech 2026: we don't build robots for show; NAO delivers real-world solutions 🦾✨ With unwavering conviction, we came to show what we build in France: a social robotics technology that is useful, easy to use, certified, highly human-centered, and brings people together.
Invasive brain-computer interfaces are moving from technology demonstrations toward early validation of commercial pathways, with a triple inflection point in regulation, clinical use, and AI decoding arriving together in 2026. China's first invasive brain-computer interface product has been approved and granted a national medical insurance billing code; US companies including Neuralink, Synchron, Paradromics, and Precision are entering multi-path clinical trial stages along different routes such as microneedles, intravascular approaches, electrode arrays, and intraoperative methods. The combination of large language models and neural decoding algorithms is pushing the valuation ceiling of brain-computer interfaces from motor control toward language and more complex human-computer interaction.
A new method for off-policy RL with diffusion: given off-policy data, we need to determine what its diffusion latent steps would be under the *current* policy (rather than the policy that collected it), which requires inverting the diffusion process on the off-policy data.
This article is an interview with Dr. Jianyuan Wang, first author of VGGT, which won the CVPR 2025 Best Paper Award, and of VGGT-Ω, a CVPR 2026 Best Paper finalist. He is currently pursuing a PhD through a joint program between the Oxford Visual Geometry Group and Meta AI.