Industry Briefing
A single destination for timely, editor-curated robotics news from around the world.
Small-AI Models Gain Traction Around the World
One morning in 2019, Adebayo Alonge was in a Cape Town hotel room, preparing to demonstrate his startup’s AI answer to a serious problem in African health care: counterfeit medication, which kills thousands of people across the continent every year.The RxScanner is a handheld spectrometer that scans a pill with infrared light, then sends the item’s molecular profile to an AI model equipped with a pharmaceutical database. In seconds, the AI identifies the medication from its molecular profile—or reports that it’s phony.Pharmacies were using the system in more than a dozen countries, including Ghana, Kenya, Myanmar, and Alonge’s native Nigeria. But that morning in South Africa, it didn’t work. “I was shocked,” Alonge says.The spectrometer connected to the AI model—but the data center was 14,000 kilometers away and bandwidth was limited. “Our server was in the United States, and just to get the result of a single scan was taking me over 5 minutes.”So Alonge immediately asked his engineers to shrink the AI model down to a smaller, low-power, unconnected version that could run entirely on his Android phone. They produced it 2 hours later, and that saved the demo.More importantly, the work birthed a new version of his device, which can authenticate a pill in places without broadband, computers, or even reliable electricity. It also turned Alonge into an advocate for this kind of “small AI.”Small AI for Global Health Care AccessSmall AI is a far cry from wealthy nations’ colossal large language models (LLMs), hyperscale data centers, multibillion-dollar investments, and debates about AI consciousness. But for millions of people around the world, the only AI that matters, and often the only kind available, is small. (According to a World Bank Report issued in November, only 0.7 percent of internet users in the world’s poorest countries have used ChatGPT, compared to a quarter of all internet users in the most developed nations.)“Most people are discussing AI from the LLM/generative side. But that needs a lot of computing power, electricity, massive data, and skilled people to manage it,” Ajay Banga, president of the World Bank, said last January at the World Economic Forum, in Davos. “Outside the developed world, other than maybe India and China, very few countries have that combination.”By contrast, small AI can deliver useful, even life-saving services to people in areas that have none of those things, Banga said. In India, where the government’s AI plans call for more development of small AI, many such systems are working for farmers.For example, a drone-based system developed by Bala Murugan and colleagues at the Vellore Institute of Technology, in India, takes photos of cashew plants and quickly identifies those with splotches that indicate disease. All the processing takes place on the drone itself, so there’s no need for a computer on-site, nor for a connection to a central server.Using small language models trained for a specific problem, and sometimes running on cheap, low-power devices, other small-AI implementations have been developed to identify ant infestations in a Uruguayan vineyard, detect the presence of malaria-carrying mosquitoes in a number of nations, and run electrocardiograms from an Arduino device in parts of Brazil that lack access to more complex equipment.“This is the most important area in AI nowadays,” says Marcelo José Rovai, a professor at the Institute of Engineering and Information Systems at the Federal University of Itajubá, in Brazil, who was involved in all three projects. “It’s growing very fast.”Low-Power, Small-AI Models on Devices Small AI models can run on a variety of low-power devices, including [from left to right] an Arduino Nano 33 BLE Sense, a Seeed Wio Terminal, and an Arduino Portenta.Moez AltayebFor Alonge, Rovai, and other advocates, small AI is not just “a promising trend,” as that November World Bank report calls it. It may be, in the long term, the form of AI that will touch the most lives and remain sustainable after some of the giant models become too costly for most users.“I think the future of AI is not like one giant model, at a center. I think it’s millions of small, precise models deployed at the edge, each one solving like a specific problem, a specific context,” Alonge says. This is partly because much of humanity—including people in parts of rich countries as well as the developing world—lives without access to cutting-edge frontier models. But, he says, it’s also because those models are not sustainable.“If someone is not subsidizing it, most people will not be able to afford those models. So those of us who are said to be small-AI developers are the ones who will have to build for the majority of the world,” Alonge says.There is no strict definition of “small AI,” but people often use the term for language models with at most a few billion parameters. (Compare that to cutting-edge models, which can include more than a trillion.) That’s small enough to run directly on a phone or a Raspberry Pi. That’s what allows these applications to run on devices without a connection to a data center and use only a few watts of power, often supplied by a battery or a solar panel.Despite their small footprint, these models aren’t fundamentally different technology from that of gigantic AI models, Rovai says. Many instances of small language models were created the same way the phone-based version of Alonge’s pharmaceuticals scanner was—by “pruning” large models, or removing the parameters that weren’t involved in the task. The result is a system that’s less capable generally but still very good at the specific job it was pruned for, Rovai says. A lighter version of RxAll’s RxScanner spectrometer sends its results to an AI model run locally on a phone to check that a drug’s molecular signature is genuine.RxAllOther small models are created by “distillation.” They are trained to mimic a large model, until their performance approaches that of their “teacher,” Rovai says. In other cases, a larger model’s precision is reduced, for example, so that a model run on 32-bit architecture can run on 8-bit designs. In situations where the machine learning application is being used to classify data or predict patterns (like an ant infestation), it’s trained from the beginning on a small device, not derived from a larger model at all. Running all these small, specialized systems is becoming easier, Rovai says, for two reasons.The first reason is that hardware is getting better and more capable while using less power, he says. This means more and more phones can run small AI—especially those equipped with neural processing units, which are specialized chips that handle AI tasks like facial recognition and changing the brightness, shadows, or contrast in a photo.In 2025, slightly more than a third of all smartphones shipped worldwide were capable of running generative AI, and that figure will reach 45 percent by the end of this year, according to the technology research firm Counterpoint. By the end of next year, slightly more than half of all smartphones will be able to run a small AI model.The second reason Rovai cites is the shrinking footprint of language models. Both Google DeepMind’s Gemma 4 (released in April) and Alibaba’s Qwen 3.5 are “fantastic” for small AI, Rovai says. Both models are “open weight,” meaning users can adjust the connections between parameters to suit their needs. This makes it easy, for example, “to take a lot of data from, say, the milk industry and retrain the model specifically on that,” Rovai says.Rovai illustrated these reasons on a Zoom call, using one of his most recent experiments. Holding up a device, he says, “This is the new Arduino UNO Q—a US $50 device with a Qualcomm chipset. I’m running a language model here, which collects data from sensors and analyzes that data to detect tiny pools of water where mosquitoes might be breeding. It takes 3 watts to run it.”Support for Small-AI DevelopmentConvinced that millions of people are already benefiting from these kinds of applications, the World Bank now actively promotes small AI with grants, mentorship programs, financing, technical advice, and models of government policies that are friendly for small-AI development. For example, in Rwanda, the World Bank is backing a government program to help low-income households get devices that can run AI.All that said, no one claims that large language models are going away entirely. To create a generative AI that can run on a phone or other small device requires the architectural insights, data processing, and results of a larger model, Rovai says. “We need the big models to create these smaller models.” And for all that small AI can benefit people without access to big AI, the technology can’t solve the larger problems of development and digital inequality, Alonge says. Implementing small AI won’t allow nations to escape the challenge of creating an ecosystem to support AI: reliable power, a supply chain that works, and an educational system that develops the talents needed to create AI tools.Though his drug-scanning system can run for days on a phone with no connection, “you still want to be able to enable periodic syncing for updates with new signatures for the medications and analytics,” Alonge says. “And even when you are using batteries, reliable power is important. That phone battery is not going to last forever.”In many parts of the world, the future of small AI isn’t assured, he says. “It works, and many places will eventually need to use it. The question is whether or not the political actors are wise enough to invest in infrastructure to support it long term.”
Small-language-models Artificial-intelligence Llms
IEEE Rolls Out Large Language Models Virtual Training Course
Large language models (LLMs) have transitioned from research labs to everyday use in engineering, significantly altering how digital infrastructures are developed and maintained. As technical professionals increasingly rely on LLMs for complex tasks—such as identifying vulnerabilities in source code and converting fragmented discussions into detailed specifications—the demand for expertise in this technology is surging. According to MarketsandMarkets, the LLM technology market is projected to grow by approximately 33% annually through 2030. To effectively utilize LLMs, engineers must move beyond basic interactions and understand the underlying transformer architecture that enables these models to process vast datasets simultaneously. This knowledge is crucial to mitigate risks associated with inaccuracies, often referred to as "hallucinations," and to ensure reliable performance in coding and data handling. Key advancements include integrating LLMs with application programming interfaces (APIs) for direct database connections, addressing hallucination issues through retrieval-augmented generation (RAG), and prioritizing data security by establishing private model instances. Additionally, LLMs automate repetitive tasks, allowing engineers to focus on higher-level design and problem-solving. To bridge the growing knowledge gap, IEEE has launched an online program titled "Large Language Models Demystified," designed to equip technical professionals with a deeper understanding of LLMs. The curriculum covers the evolution of AI technology, transformer architectures, and practical model-building exercises. Participants will earn professional development credits and a digital badge upon completion, enhancing their credentials in this rapidly evolving field. Organizations interested in training their teams can consult with IEEE for tailored enrollment options.
Ai Type-ti Education Ieee-educational-activities Large-language-models Ieee-products-and-services
Liquid AI's smallest model yet LFM2.5-230M beats models 4X its size at data extraction, can run 'anywhere'
Liquid AI, a company founded by former MIT computer scientists, has unveiled its latest AI language model, LFM2.5-230M, which is designed for efficient data extraction and local deployment on devices such as smartphones and laptops. Released today, this 230-million-parameter model is noted for its ability to run on various hardware platforms, outperforming larger models like Alibaba's Qwen3.5 and Google's Gemma 3 in specific benchmarks. Targeting developers and engineers, LFM2.5-230M operates under a dual-use commercial license, allowing free access for individuals and companies with annual revenues below $10 million, while larger enterprises must secure a paid agreement. The model distinguishes itself by utilizing the LFM2 architecture, enabling high inference speeds with a minimal memory footprint, making it suitable for edge computing. Liquid AI's launch reflects a broader industry shift towards architectural efficiency rather than sheer parameter counts, as major AI firms focus on models with hundreds of billions of parameters. The LFM2.5-230M is specifically tailored for lightweight data extraction tasks, allowing businesses to automate processes without relying on costly cloud services. In practical applications, the model has been successfully deployed in a humanoid robot, demonstrating its capability to process complex commands efficiently. Available immediately on platforms like Hugging Face, LFM2.5-230M aims to revolutionize how enterprises manage data extraction, moving away from traditional, rigid systems to more adaptable AI-driven solutions.
Technology
Why Are Large Language Models So Terrible at Video Games?
Recent advancements in large language models (LLMs) have led to significant improvements in various domains, particularly in coding. However, a notable limitation remains: LLMs struggle to play video games effectively. Despite some successes, such as Gemini 2.5 Pro defeating Pokémon Blue in May 2025, these models often perform poorly compared to human players, making frequent mistakes and requiring specialized software to assist them. Julian Togelius, director of New York University’s Game Innovation Lab and co-founder of AI game-testing firm Modl.ai, discussed these challenges in a recent interview with IEEE Spectrum. He highlighted that while coding resembles a well-structured game with clear tasks and immediate feedback, video games present a more complex landscape that LLMs have yet to navigate successfully. Unlike games like chess or Go, which have been mastered by AI through retraining, video games vary significantly in mechanics and input requirements, complicating the development of a general game AI. Togelius pointed out that the lack of comprehensive benchmarks for video games further hinders LLMs' performance. While benchmarks have driven improvements in coding, the diverse nature of video games makes it difficult to establish similar metrics. He noted that current LLMs perform poorly even compared to basic algorithms in gaming contexts, primarily due to insufficient training data and challenges in spatial reasoning. Despite their coding capabilities, LLMs cannot engage in the iterative process of game development, which involves testing and refining gameplay. This disparity raises questions about the future of AI in mastering video games and its implications for broader AI applications.
Llms Artificial-intelligence Video-games
Visual Language Models Train Robots to Read Human Emotions
A recent study led by Seung Chan Hong at the University of Melbourne explores the emotional capabilities of collaborative robots as they increasingly work alongside humans. Published on May 18 in IEEE Robotics and Automation Letters, the research investigates how robots can better understand human emotions through contextual cues, beyond just facial expressions. Involving 40 volunteers, the study trained a vision language model (VLM) to interpret emotions based on video interactions where robots handed objects to humans. The VLM outperformed traditional AI systems, scoring 0.86 in emotional accuracy compared to 0.77 for conventional methods. This improvement is attributed to the VLM's ability to consider the entire context of interactions rather than isolated facial expressions. In a follow-up experiment, participants interacted with a robot that was programmed to make an error, receiving either an emotionally adaptive apology or a standard one. The majority preferred the adaptive response, but trust in the robot diminished after it failed to complete its task, highlighting that emotional responses cannot compensate for a lack of functionality. While the VLM effectively recognized emotions from a third-party perspective, its accuracy dropped when compared to participants' self-reported feelings, indicating that robots still struggle to fully understand human emotions. The findings suggest that while emotional adaptivity is valuable, the primary concern for users remains the robot's competence in performing tasks.
Robotics Journal-watch Ai-models Emotion-recognition
When Machines Learn to See Like Experts: The Rise of Vision Language Models in Manufacturing
In a significant development for the manufacturing sector, experts have highlighted the transformative potential of Variational Latent Models (VLMs) in enhancing quality assurance processes. While acknowledging that VLMs will not address every challenge faced in the realm of artificial intelligence within manufacturing, they emphasize that these models provide a unique capability that surpasses existing technologies, particularly in high-complexity production environments. This advancement comes at a time when industries are increasingly seeking innovative solutions to improve efficiency and accuracy in their operations. As manufacturers strive to meet rising demands and maintain high standards, the adoption of VLMs could represent a pivotal shift in how quality assurance is approached, ultimately leading to more reliable and efficient production outcomes.
LA4VLA: Redefining VLA Pretraining for Robots to Learn Action Language
Researchers have introduced the LA4VLA framework, a new approach that enhances the capabilities of robots in understanding language commands and executing actions. This framework distinguishes language-action supervision from visual input, enabling robots to learn the relationship between commands and actions independently of visual cues. The study, which highlights the limitations of traditional Vision-Language-Action models, was conducted to address the tendency of these models to rely on visual inputs when confronted with conflicting information. By focusing on a more robust language-action learning process, the LA4VLA framework aims to improve the overall understanding of how language influences robotic actions.
Vision-Language-Action Robotics Machine Learning AI Training
ConlangCrafter Turns AI to Imagining Languages
Researchers from the University of California, Berkeley, Carnegie Mellon University, and Tel Aviv University have developed an AI model named ConlangCrafter, capable of generating new languages. The findings, published on June 27 in the Proceedings of the Association of Computer Linguists, highlight ConlangCrafter's ability to create diverse and rule-abiding languages, surpassing traditional human efforts in language construction. Led by linguist Gašper Beguš, the team designed ConlangCrafter to apply various linguistic rules, including phonology and morphosyntax, while incorporating a random number generator to ensure each language is unique. The model can even simulate unconventional communication systems, such as a hypothetical language for cephalopods that utilizes colors and gestures. The researchers evaluated the generated languages for diversity and consistency, finding that ConlangCrafter produced languages that were twice as diverse and 70% more consistent than those created by general-purpose language models. This advancement could aid natural language processing researchers in understanding how language structure impacts model performance. While ConlangCrafter is currently available for free online, it has limitations in more complex linguistic areas like semantics and contextual usage. Beguš envisions future research exploring the Sapir-Whorf hypothesis, which posits that language influences thought and perception, potentially leading to simulations of societies with distinct languages.
Llms Artificial-intelligence Languages
How World Models and VLA Can Be Implemented: Insights from Top Experts in Embodied Intelligence
At the 2026 Zhangjiang Embodied Intelligence Supply Chain Conference, a roundtable discussion brought together leading experts in robotics to explore the critical role of world models in embodied intelligence. The event highlighted various industry challenges, particularly the necessity for robust data infrastructure and the integration of visual-language-action models with world models. Experts emphasized that high-quality data and innovative technological solutions are essential for advancing the field. The conference served as a platform for addressing these pressing issues, aiming to foster collaboration and drive progress in robotics and artificial intelligence.
Embodied Intelligence World Models Robotics Data Infrastructure AI Integration
BABA Stock: The Physical AI Race Heats Up as Alibaba Releases New AI Models for Robots
Alibaba Group Holding Limited has intensified its efforts in the rapidly evolving artificial intelligence (AI) sector by launching the Qwen Robot Suite, a new collection of AI models designed for robotic applications. Announced on June 21, 2026, the suite aims to enhance robots' capabilities in understanding their environments, navigating complex spaces, and executing tasks based on natural language commands. Developed by Alibaba’s Tongyi Lab, these models are currently being tested with select customers of Alibaba Cloud, marking the company's strategic move into the burgeoning Physical AI market. This initiative comes as Alibaba seeks to establish itself as a significant player in the multi-trillion-dollar robotics industry amid increasing competition. Despite facing challenges in its stock performance, with shares down 44.8% from a 52-week high, Alibaba reported a 3% year-over-year revenue growth for the fiscal fourth quarter, driven by a 38% surge in Cloud Intelligence revenue. However, the company also experienced a sharp decline in profitability, with non-GAAP net income plummeting to $12 million from nearly $30 billion in the previous year. As Alibaba continues to invest heavily in AI infrastructure and cloud capabilities, the launch of the Qwen Robot Suite reflects its commitment to innovation in the face of market pressures and evolving consumer demands.
Alibaba unveils Qwen-Robot series with three foundation models for embodied AI
On Tuesday, the Qwen team unveiled a new robotics suite that includes three foundational models: Qwen-RobotNav, Qwen-RobotManip, and Qwen-RobotWorld. These models are designed to integrate language processing with various physical actions, enhancing the capabilities of mobile robotics. Qwen-RobotNav, in particular, advances vision-language integration by employing controllable observation encoding and tool-based interfaces. This innovative model consolidates four essential tasks into a single framework, which includes instruction following and goal-directed navigation. The release aims to improve the interaction between language and robotics, paving the way for more sophisticated and versatile robotic applications.
News Feed
AI’s next frontier, world models, and why China is ahead of the pack
In the rapidly evolving field of artificial intelligence, world models that simulate physical environments are gaining attention as the next frontier, surpassing traditional large language models. Recent developments indicate that China is leading the way in this area, outpacing the United States in the deployment of these advanced systems. These world models, which comprehend the physical laws governing the universe, are already being utilized to enhance AI applications, including robotics and autonomous vehicles. As of October 2023, China has integrated these technologies more extensively than its American counterparts, marking a significant advancement in the global AI landscape. This trend highlights the growing competition between the two nations in harnessing AI's potential for practical applications.
Interview with Wang Zhongyuan: VLA will survive, but world models are the future.
In recent months, the concept of "World Model" has gained significant traction within the AI and robotics sectors, driven by underlying industry anxieties. As AI technology has rapidly evolved over the past two years, limitations in embodied intelligence have become apparent, revealing that while robots can recognize objects, they struggle to understand physical interactions and causal relationships. The World Model aims to bridge this gap by enabling robots to learn the laws of the physical world. At the forefront of this exploration is Wang Zhongyuan, the director of the Beijing Academy of Artificial Intelligence, who identifies four distinct paths in the development of World Models. These include language-centered models, pixel-centered models, 3D structure-centered models, and visual representation-centered models. The Beijing Academy is pioneering a fifth approach that integrates language and visual data into a unified latent space representation, allowing for more complex interactions and predictions. Wang emphasizes that the World Model's potential lies in its ability to enhance embodied intelligence, enabling robots to understand and predict physical interactions over time. He envisions a future where World Models serve as the foundational brain for robots, capable of complex reasoning and decision-making in real-world scenarios. However, he cautions that achieving this goal will require significant advancements in data collection and model training, with a timeline of three to five years anticipated for substantial progress. As the field continues to evolve, the competition will focus on the ability to create models that accurately reflect the complexities of the physical world.
Cambridge Team Develops AI System to Generate 10,000 3D Models in One Day
A collaborative research team from Cambridge and Oxford has developed an innovative AI system named Articraft, which can produce more than 10,000 interactive 3D models within a 24-hour timeframe. This groundbreaking technology employs a specialized software development kit (SDK) that enables large language models to generate code for 3D object creation directly, eliminating the need for conventional modeling software. This advancement not only enhances efficiency but also significantly reduces costs associated with 3D modeling.
3D Modeling AI Technology Machine Learning Robotics
NVIDIA Launches Nemotron 3 Nano Omni Model, Unifying Vision, Audio and Language for up to 9x More Efficient AI Agents
NVIDIA has introduced the Nemotron 3 Nano Omni, an innovative open multimodal AI model designed to enhance the efficiency of AI agent systems. Announced today, this model integrates vision, speech, and language capabilities into a single framework, addressing the common issue of time and context loss that occurs when data is transferred between separate models. By streamlining these processes, the Nemotron 3 Nano Omni aims to improve the performance of AI applications across various domains. This advancement is particularly significant as it allows for more cohesive and contextually aware interactions, marking a notable step forward in the development of AI technologies.
Mimic Robotics Open-Sources "mimic-video" Recipe to Accelerate Video-Action Models
Mimic Robotics, a company based in Zurich, has unveiled its innovative "pixel-to-action" architecture, which is designed to transform the current landscape of artificial intelligence by moving away from traditional static vision-language models. This release, which includes both the code and accompanying research, marks a significant shift towards utilizing dynamic video-based foundations. The initiative aims to enhance the capabilities of AI systems, enabling them to better interpret and respond to visual information in real-time. By sharing this technology, Mimic Robotics seeks to foster advancements in the field and encourage further exploration of video-based AI applications.
Mimic Robotics Europe open-source ETH Zurich
DeepMind CEO Demis Hassabis: World Models and 'Infinite Training Loops' are the Keys to AGI
In the season finale of the Google DeepMind podcast, Demis Hassabis discussed the limitations of language models in advancing robotics. He emphasized that while language models play a crucial role, they are insufficient on their own for the development of physical AI. Hassabis highlighted the importance of integrating world-generators, such as Genie, with agents like SIMA to create a more effective synergy that can enhance robotic capabilities. This collaboration aims to address the challenges faced in the field of AI, particularly in bridging the gap between virtual understanding and real-world application. The insights shared during this episode reflect ongoing efforts to innovate and improve the functionality of AI in practical settings.
DeepMind Google embodied-ai
These new Roombas are smaller and cheaper
iRobot has unveiled a new range of Roomba robot vacuums, enhancing its existing lineup just a year after launching its first lidar-equipped models. This week, the company introduced eight updated vacuums, which boast improved suction power and a more compact design for better navigation. Notably, these new models are also more affordable, with some priced up to £200 (approximately $270) less than their predecessors. The introduction of these cost-effective and technologically advanced vacuums aims to attract a broader customer base and solidify iRobot's position in the competitive home cleaning market.
News Robot Smart Home Tech
Emily Bender Sets the Record Straight on “Stochastic Parrots”
In March 2021, a notable paper titled “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” was published by a team of four linguists and computer scientists, including Timnit Gebru and Margaret Mitchell, shortly after their controversial dismissal from Google. The paper critiques large language models, suggesting they generate text through statistical predictions rather than genuine understanding, coining the term "stochastic parrot" to illustrate this concept. As the analogy gained traction beyond academia, it sparked debates and inspired projects, including a shoulder-mounted robot named the Stochastic Parrot. On the five-year anniversary of the paper, lead author Emily M. Bender, a professor at the University of Washington, addressed common misconceptions surrounding the term in a recent blog post and an interview with IEEE Spectrum. Bender emphasized that the phrase specifically refers to large language models and not to all forms of artificial intelligence, which she believes oversimplifies the technology and complicates discussions about its implications. She highlighted the importance of clear terminology in understanding and regulating technology, noting that many discussions conflate different AI applications, such as chatbots and protein folding algorithms. Bender also acknowledged that the paper overlooked significant issues, such as exploitative labor practices in data collection, which she now believes should have been included. The ongoing discourse around language models continues to evolve, reflecting the complexities of artificial intelligence and its societal impact.
Emily-bender Large-language-models Llms Ai-ethics
36Kr Exclusive: Four Key Propositions for ByteDance's AI by 2026
ByteDance is setting ambitious goals for its AI initiatives in 2026, focusing on four key areas. The company aims to enhance world model training, targeting performance levels comparable to Google's leading model, Genie 3, by the end of the year. Additionally, ByteDance plans to maintain its leadership in video models while exploring new avenues like dynamic generation. The company is also committed to strengthening its coding capabilities, emphasizing the importance of data feedback and evaluation to improve agent performance, particularly in office applications. Despite recent advancements, including the launch of Seed 2.0 and Seedance 2.0, ByteDance faces challenges in the world model arena, having entered the field later than competitors. The company established a research group in 2025 to explore visual-language-action models and has since set a clear goal for world model development. However, internal assessments indicate that performance still lags behind global standards by approximately 10%. In parallel, ByteDance is accelerating the commercialization of its Doubao platform, which has seen a surge in daily active users, reaching 200 million. The company plans to introduce paid features and enhance its offerings for professional users, particularly in sectors like finance and law. Doubao's strategy includes differentiating itself in the crowded AI tools market and expanding its presence internationally, with a focus on small language markets. As ByteDance navigates these challenges, it aims to leverage its engineering expertise and data resources to emerge as a leader in the evolving AI landscape.
Startup Wants to Run AI Inference From Space
The rapid growth of large language models is driving a global surge in energy demand for data centers, prompting operators to seek alternative power sources. Among them is Orbital Inc., a Los Angeles-based startup that recently emerged from stealth mode to announce plans for space-based data centers. Backed by venture capital firm Andreessen Horowitz, Orbital aims to utilize solar energy from a constellation of small satellites in low Earth orbit to power AI inference workloads, such as chatbots. Orbital's founder and CEO, Euwyn Poon, emphasizes the limitations of terrestrial energy sources, stating, “There simply isn’t enough capacity here [on Earth], and the only way is up.” The company envisions a network of up to 10,000 satellites, each equipped with GPU server racks powered by solar panels. The first test of this concept is scheduled for 2027, with a prototype satellite launch aboard a SpaceX Falcon 9 rocket. While Orbital's approach aims to reduce launch costs and improve efficiency, it faces significant engineering challenges, including radiation effects on GPUs, thermal management in space, and maintenance difficulties. Experts like Dr. Amit Verma from Texas A&M University caution that the operational feasibility of such systems will depend on the specific applications they support. Despite these hurdles, Orbital plans to finalize its satellite designs by 2026 and establish a manufacturing facility by 2028, with the goal of tapping into major AI firms as customers. Poon remains optimistic about overcoming technical challenges, asserting that their engineering efforts will pave the way for the future of space-based data processing.
Data-center Space Ai Inferencing
NVIDIA RTX Accelerates 4K AI Video Generation on PC With LTX-2 and ComfyUI Upgrades
In 2025, advancements in artificial intelligence on personal computers reached a significant milestone, with small language models (SLMs) achieving nearly double the accuracy compared to the previous year. This improvement has notably narrowed the performance gap between these PC-class models and larger, cloud-based language models (LLMs). The surge in AI development is attributed to enhanced developer tools and techniques that have emerged, enabling more efficient training and deployment of SLMs. As a result, users can now access more powerful AI capabilities directly on their PCs, marking a pivotal shift in the landscape of AI technology.
Māori Text-to-Speech Model Spurns Big Tech’s Values
Researchers at the University of Waikato in New Zealand have developed a high-fidelity synthetic voice for te reo Māori, the indigenous language of the country, in response to concerns over the ownership and control of Māori language data by foreign technology companies. Led by associate professor Te Taka Keegan and his former master's student Kingsley Eng, the project was motivated by a desire for "sovereign digital systems" that prioritize Māori ownership of their language resources. The initiative began with the recording of 4.5 hours of data from Ngaringi Katipa, a fluent speaker and language mentor, which was later expanded to 7 hours and 45 minutes. The researchers faced challenges due to the unique linguistic features of te reo Māori, such as vowel length and digraphs, which can alter meanings. They employed a phoneme-based approach to training the text-to-speech model, utilizing open-source tools and testing various neural architectures to achieve an effective AI voice with a word error rate of 6.78 percent. Despite receiving funding from Google, Keegan emphasized that the ownership of the voice model remains a collective responsibility of the Māori community, particularly the tribes affiliated with Katipa. The project aims to empower Māori language speakers and establish a framework for similar initiatives among other indigenous communities globally. Keegan envisions a future where community-owned language models can preserve and promote indigenous knowledge, ensuring that technology serves to empower rather than diminish cultural heritage.
Artificial-intelligence Languages Ai-models
Better Hardware Could Turn Zeros into AI Heroes
Researchers at Stanford University have developed a groundbreaking hardware accelerator named Onyx, designed to enhance the efficiency of artificial intelligence (AI) computations by leveraging the concept of sparsity. This innovation comes in response to the growing energy demands and carbon footprint associated with increasingly large language models (LLMs), such as Meta's recent Llama release, which boasts 2 trillion parameters. Onyx aims to address the limitations of current hardware, which often fails to fully utilize the sparse nature of AI models, where many parameters are effectively zero. By re-engineering the architecture to support both sparse and dense computations, Onyx achieves significant energy savings—consuming up to one-seventieth the energy of traditional CPUs and performing computations eight times faster on average. The development of Onyx reflects a broader trend in AI research, where experts are exploring new algorithms and hardware solutions to mitigate the environmental impact of AI technologies. The team at Stanford plans to expand Onyx's capabilities to support a wider range of AI operations, potentially revolutionizing the field and paving the way for more sustainable AI practices. As the demand for efficient AI solutions grows, Onyx represents a promising step toward balancing performance and energy consumption in machine learning.
Ai-models Gpus Energy-efficiency Data-compression
Can AI Chatbots Reason Like Doctors?
A recent study published on April 30 in the journal Science reveals that OpenAI's large language model (LLM) has outperformed physicians in clinical reasoning tasks using real emergency room records. This research comes amid growing scrutiny of the reliability of medical information provided by chatbots, with some studies highlighting impressive diagnostic capabilities while others point to inaccuracies and fabricated information. OpenAI has introduced tools like ChatGPT for Clinicians and ChatGPT for Healthcare, aiming to assist medical professionals. The study involved comparing the performance of the LLM with that of physicians during various stages of emergency care, demonstrating that the AI model consistently provided accurate or close diagnoses more frequently than human doctors. Despite the promising results, researchers, including coauthor Arjun Manrai from Harvard Medical School, caution against interpreting these findings as a signal that AI could replace doctors. Instead, they emphasize the need for further research and clinical trials to explore how LLMs can be effectively integrated into medical practice. Experts like Mickael Tordjman from the Icahn School of Medicine stress the importance of developing reliable evaluation methods for LLMs in clinical settings. As the technology evolves rapidly, there is an urgent need to address regulatory and liability questions surrounding its use in healthcare. While acknowledging the potential benefits of AI in medicine, researchers advocate for responsible innovation and careful evaluation to ensure patient safety and effective integration into clinical workflows.
Large-language-models Llms Chatbots Medical-ai Ai-safety Openai
Tsinghua alumni launch AI startup, securing hundreds of millions in angel funding for the automotive industry.
Embodied intelligence company Guangxiang Technology has successfully secured hundreds of millions in angel funding, with significant participation from leading investors including Zhuhai Technology Industry Group, Xingsheng Capital, and several others. Founded in April 2025 through a collaboration between Tsinghua University's Vehicle and Transportation Institute and its AI Institute, Guangxiang aims to advance the development of its physical native base model and commercialize embodied intelligent robots for industrial applications. The company's founder and CEO, Zhang Tao, previously led the spatial perception engine at Amap, while co-founder Professor Li Shengbo is a renowned expert in reinforcement learning and autonomous driving. The team, which includes members from major tech firms like Alibaba and Huawei, is focused on a unique technological approach that diverges from mainstream visual-language-action models. Guangxiang's flagship product, the Phi-Bot X1, was launched in June 2026 and is designed for industrial environments. The robot has demonstrated impressive capabilities, completing a full welding operation on an automotive production line without errors during a 21.5-hour continuous run at the 2026 ATC exhibition. The company plans to expand its applications in the automotive sector, targeting the remaining 30% of automation gaps that traditional methods cannot address. Zhang envisions a robust market for automotive production line robots, estimating a potential market size of around 100 billion yuan in China. Guangxiang Technology is committed to refining its products and scaling operations, with a focus on real-world applications and continuous feedback to enhance its offerings.
Context is king: How Avride uses cloud VLMs as a safety net for delivery robots
Avride is enhancing the environmental awareness of its delivery robots by utilizing vision-language models (VLMs). This innovative approach aims to improve the safety and efficiency of the robots as they navigate various environments. By integrating cloud-based VLM technology, Avride is able to provide its robots with a better understanding of their surroundings, enabling them to respond more effectively to dynamic conditions. The implementation of these advanced models is part of Avride's ongoing efforts to ensure safer delivery operations and to adapt to the complexities of real-world environments.
Artificial Intelligence Artificial Intelligence / Cognition Autonomous Mobile Robots (AMRs) Logistics Mobility / Navigation News
Japan Pioneered Humanoid Robots—Can It Now Catch China?
“In the future, the relationship between humans and robots will deepen, and the distinction between them will probably disappear.” This prediction, from one of the attendees at the recent Humanoids Summit in Tokyo, might have been unremarkable had it not come directly from an android that was first introduced to the world 20 years ago. Geminoid HI-6 is the sixth-generation of a robot originally designed in 2006. The mechanical twin of Osaka University professor Hiroshi Ishiguro, Geminoid HI-6 is now equipped with a large language model trained on Ishiguro’s own writings and interviews. It has advanced conversational skills and can even have a chat with its creator, an eerie spectacle. But at the Humanoids Summit, Geminoid was one of the few humanoid robots from Japan, the country that pioneered the form factor.While the event in Tokyo only had about 40 robots on display, Chinese systems outnumbered Japanese by roughly three to one. Some Japanese robotics firms were even using Chinese robots in their own technology demonstrations, something that would have been unthinkable in the recent past—one Japanese engineer described the situation as “sad.” The conference was a stark reminder of how Japan has ceded its early lead in humanoid robot development to overseas competitors, and the challenge it now faces to secure a place in an ecosystem increasingly dominated by general-purpose robots powered by AI. Twenty-five years ago, Japan was turning out groundbreaking humanoids that were showstopping in their abilities, but they were not commercialized as practical machines in any meaningful way. Heavily influenced by science fiction and lacking practical applications, they were mostly expensive technology demonstrations that were eventually mothballed. What Japan retains, however, is robotics design and know-how, which it must leverage to be a key player in the rapidly evolving humanoid ecosystem. Learning to Walk—Then Standing StillTo anyone who has seen recent videos of Chinese humanoids doing kung-fu and synchronized acrobatics, as well as half-marathon races, China’s remarkable progress in the field is nothing new. At the Humanoids Summit, Toyota showed a video of its latest basketball-playing robot, and Honda exhibited its latest robot hand, but the full-scale humanoids on the floor were mostly Chinese–the kid-size K1 machines from Booster Robotics of Beijing were dancing to Michael Jackson tunes. The full-scale G1 humanoid from Unitree Robotics of Hangzhou was also doing demos. “You cannot sell these bipedal systems in Japan for safety and compliance reasons,” says Shuichi Nagao, a frequent visitor to China as CTO of Omakase Robotics, a division of Zeals, a Japanese humanoid robot developer. Omakase was exhibiting a G1 modified with an external PC controller, a dextrous hand, a suction-cup manipulator and a sensor “hat” with an extra speaker, mic and camera. “In China, the government is pushing humanoid development. They didn’t have an industry 20 years ago. The people pushing it are young, in their 20s and 30s. It’s a really different mentality out there,” says Nagao. “Big players in Japan are still looking for use cases for humanoids. In China, they’re already doing mass production and reducing the cost, so other countries can’t compete with them anymore.”Another Japanese company showing off G1 bots was summit sponsor GMO AI & Robotics, a subsidiary of Japanese internet company GMO. It’s using the robots in partnership with Japan Airlines to load and unload cargo containers at Tokyo’s Haneda airport. The cargo project is a trial—like many other humanoid experiments—but the fact that Chinese machines have penetrated so far into Japan’s ecosystem upends a long history. In 1973, scientists at Waseda University in Tokyo built WABOT-1, considered the first full-scale humanoid robot and capable of slow bipedal locomotion, grasping objects and simple communication. It inspired Honda’s groundbreaking Asimo humanoid, but it was never commercialized. Asimo was eventually retired in 2022, the year ChatGPT was released. Two years later, Unitree’s G1 went on sale for US $16,000. China’s High Torque Technology Co. showed off its Mini Pi biped, customized with an anime-inspired head, at Humanoids Summit in Tokyo. The regular version is priced at $3,500. Tim HornyakSupply and DemandJapan’s development of humanoids happened before practical applications or widespread demand were in place, but bad timing is only part of the story—Japan also has a history of developing technologies that might appeal to domestic consumers but not necessarily those overseas. For example, decades after they first appeared, its highly engineered, multifunction toilets have only recently found a following abroad. Japan’s humanoid prowess was partly built on the back of its legendary industrial automation, yet even that stronghold has eroded. Ani Kelkar, a partner from McKinsey & Company in Boston who produces analytical reports about the robotics industry, told the summit audience that while Japan occupied the top spot in the world in manufacturing robot density (the number of multipurpose industrial robots in operation per 10,000 employees) from at least 1994 to 2009, it then slipped to second in 2014, third in 2019 and fifth in 2024. In that year, South Korea was at the top of the leaderboard with a robot density of 1,220 compared to Japan’s 446. The International Federation of Robotics estimates China now has the most operational industrial robots in the world, with around 2 million total units, approximately 4.5 times more than Japan. “The annual installation numbers are impressive too: 54 percent of all robots installed worldwide in 2024 were deployed in China,” the IFR said in a release in April 2026. “I think the loss of Japanese leadership is more to do with the rise of China as a manufacturing powerhouse including for sectors that Japan had high export levels,” Kelkar said in an email interview. “The recovery has not yet happened as Japan ‘missed’ the rapid acceleration in AI for robotics and is now playing catchup.”How Japan Can Adapt Kelkar believes Japan has a US $100 billion opportunity in general-purpose robotics, which are machines that can perform a wide variety of tasks, and it cannot rely on the slower-growing industrial robot market, which is centered on factory machines that do one simple and predictable task like welding car parts. He points to a McKinsey white paper suggesting that while Japan has much of the hardware and technology experience needed to support general purpose robot development, it must change its strategy to capture more share in AI, software, data collection and robotics platforms.Tetsuya Ogata is a professor of engineering and director of the Institute for AI and Robotics at Waseda University, the birthplace of humanoids in Japan. He briefed the summit on how a nonprofit he chairs, the AI Robot Association (AIRoA), is working with Toyota and other members to develop foundational technologies for collaborative use. For instance, AIRoA has collected some 80,000 hours of data on remote operation of mobile manipulators, and Ogata believes it’s the largest dataset of its kind. Using the data, it built and verified Vision-Language-Action (VLA) models, and it has also started data collection for dual-arm mobile manipulation. In an interview, Ogata acknowledged Japan’s struggle to find its place in the changing landscape. “The world of AI is inherently a game of scale,” says Ogata. “Therefore, Japan’s absolute prerequisite is to secure a competitive baseline of scale—in data, computing resources, and talent. Beyond that, what I consider most critical is a mindset shift: rather than trying to hoard scale within a single nation or company, we must grow stronger by collaborating with a diverse ecosystem of domestic and international players.” Specifically, this means creating a ‘collaborative domain’ to address data—the single biggest bottleneck—through industry-wide cooperation rather than data-siloing. By collectively nurturing a pre-competitive, shared data infrastructure and foundation model, individual companies can then compete on top of it with their own applications. “By offering this open ‘data ecosystem’ to the world, we can engage global players and establish a ‘third pole’ alongside the US and China,” says Ogata. “I believe this is how Japan can reclaim its global presence.”In 1999, Japan introduced the world’s first mobile internet services platform. But being first didn’t turn Japan into a smartphone manufacturing or design center—it’s now merely a supplier of parts to other countries who are leading the smartphone industry. If Japan can avoid a repeat of that experience and successfully deregulate, diversity, and commercialize its original humanoid dreams, it stands a better chance of influencing the direction of the industry and reaping billions in value. As automobiles and electronics were pillars of Japan’s industrial strategy in the last century, Japan could make humanoid robots one of its key value generators in the 21st century, an approach that would not only deliver economic benefits but give Japan greater clout in how the industry will evolve. Just like Japanese cars, electronics, and even toilets, Japanese humanoids could stand for craftsmanship and reliability. It’s a legacy that Japan can’t afford to give up.
Japan Robotics Humanoids Humanoid-robots
Chinese AI company offers a new solution for physical AI in the uncertain trillion-dollar market.
In 2026, the field of physical AI is set to emerge as a transformative force, following a consensus reached by industry leaders at the CES in Las Vegas, where NVIDIA's CEO Jensen Huang heralded the arrival of "physical AI's ChatGPT moment." Over the past two years, significant advancements have been made in five key areas: brain models, imagination engines, training environments, ontology, and commercial ecosystems, laying the groundwork for real-world applications. In the first half of 2026, global investment in physical AI surged, with over $6.4 billion raised in just the first quarter, including notable funding rounds from AMI Labs and World Labs. The industry is witnessing a clear technological divergence, with three primary paths emerging: Visual Language Models (VLM), Visual Language Action (VLA), and world models. The anticipated future architecture for physical AI is expected to integrate VLA's decision-making capabilities with world models' predictive simulations. Despite the rapid growth, the competitive landscape remains uncertain, with various companies pursuing different strategies, including those focusing solely on VLA or world models, and others exploring hybrid approaches. The ultimate goal is to develop AI that can effectively navigate and understand the complexities of the physical world, moving beyond mere reactive capabilities to proactive, autonomous decision-making. As the physical AI market is projected to expand significantly, reaching an estimated $3.26 trillion by 2040, the industry faces the challenge of ensuring that technology translates into tangible business value. Companies like Om AI are pioneering innovative models that prioritize continuous perception and spatial understanding, aiming to redefine how AI interacts with its environment. The ongoing evolution of physical AI emphasizes the importance of real-world applications and the need for AI systems that can adapt and respond to dynamic physical spaces.
The Lab Mistake That Might Revolutionize Computing
Researchers have made a significant breakthrough in artificial intelligence technology by discovering a new way to create electronic components that mimic the behavior of biological neurons and synapses. This development, which occurred in a laboratory in 2024, could drastically reduce the energy consumption associated with AI applications. Currently, AI systems rely on powerful GPUs housed in data centers, consuming up to 1,000 watts each, which is comparable to household appliances. In contrast, the human brain operates at a fraction of that energy efficiency. The team, led by researchers Mario Lanza and Sebastian Pazos, stumbled upon this innovation while experimenting with metal-oxide-semiconductor field-effect transistors (MOSFETs). They found that by manipulating the bulk terminal of a MOSFET, they could replicate neuron-like behavior, producing sharp current spikes similar to those of biological neurons. This discovery not only allows for the creation of artificial neurons but also enables the development of artificial synapses, leading to a new type of neurosynaptic random-access memory (NSRAM). The implications of this technology are vast, as it could lead to brain-inspired microchips that are more energy-efficient than current GPUs, particularly for smaller-scale AI tasks. The researchers are now focused on refining their models and conducting further simulations to optimize performance. If successful, this innovation could pave the way for a new generation of AI systems that are both powerful and environmentally sustainable.
Neuromorphic-computing Cmos Mosfet Synapse
What lessons is Ideal learning to catch up with FSD V14?
The competitive landscape of the intelligent driving industry has undergone significant changes in recent years, shifting from hardware specifications to advanced model development. Companies are increasingly recognizing that merely having larger models is insufficient for achieving generational advantages; instead, the integration of models, data, computing power, and chips into a continuous iterative loop is becoming crucial. This realization has prompted many automakers to invest in in-house research and development. Tesla has established a comprehensive ecosystem that spans data collection, training infrastructure, and self-developed chips, while Chinese companies like Li Auto, Xpeng, and NIO are also deepening their technological foundations. Li Auto has introduced its self-developed Mach M100 chip in its L8 and L9 models, which it views as a significant advancement in AI technology. In a recent discussion with Li Auto's autonomous driving and chip leaders, they emphasized that the industry should focus on the practical problems these investments aim to solve rather than merely the existence of in-house development. They outlined their strategies to achieve performance comparable to Tesla's Full Self-Driving (FSD) system, highlighting the importance of safety, efficiency, and comfort in user experience. As the industry moves towards higher levels of autonomy, the integration of vision and language models is seen as essential for developing systems that can handle complex, unforeseen scenarios. The executives noted that achieving higher levels of autonomy (L3 and L4) requires models that can reason and think like humans, underscoring the growing significance of language in AI systems. Overall, the conversation revealed the industry's focus on enhancing AI capabilities through innovative chip design and data utilization, aiming for a future where autonomous driving technology can meet the challenges of real-world driving conditions.
RoboScience launches Visics, a versatile embodied model for cross-ontology, cross-object, and cross-task applications.
On June 24, RoboScience, a company specializing in embodied intelligence, unveiled its self-developed Visics large model, introducing the innovative VLOA (Vision-Language-Object-Action) architecture. This announcement marks a significant advancement in the field, demonstrating the model's applications in real-world scenarios such as furniture assembly, dexterous grasping, and dynamic assembly lines. The current landscape of embodied intelligence lacks a universally accepted foundational representation unit, which hampers data collection, model learning, and the transfer of knowledge to new contexts. Traditionally, models have focused on replicating specific robotic movements tied to particular tasks, limiting their adaptability to new robots, objects, or environments. Founder and CEO Tian Ye highlighted three major challenges in robotic operations: poor generalization, difficulty in precise manipulation, and cumulative errors in long-range tasks. To address these issues, RoboScience has developed a new foundational representation unit from the ground up. The Visics model employs a dual-engine architecture, consisting of an embodied world model and a universal operation model, each operating independently. The embodied world model utilizes vast amounts of internet video data to learn the physical dynamics of objects, while the operation model translates object trajectories into actionable commands for robots. This layered design enhances the model's generalization capabilities across various robotic platforms and tasks. RoboScience's innovative approach also includes a high-precision simulation engine, RoboMirage, which, combined with automated video data annotation, significantly reduces data acquisition costs. The company aims to build a comprehensive dataset of over 1 terabyte of high-quality manipulation trajectories by 2026. Since its inception, RoboScience has garnered support from multiple investors and established research and production centers in major Chinese cities. The company plans to collaborate with various sectors, including retail and logistics, to standardize robotic products for industrial and commercial applications by the end of this year.
Interview with Digid’s Nils Könne and Christian Kreil: Nanoscale sensors could help solve robotics’ tactile sensing challenge
The swift advancement of artificial intelligence and robotics is drawing significant attention to software and powerful processors, particularly large language models. However, experts emphasize that for robots to function effectively in real-world settings, they require a fundamental capability: advanced environmental sensing and understanding. This necessity is driving increased interest and investment in cutting-edge sensing technologies, as researchers and developers seek to enhance robots' interaction with their surroundings. The push for these innovations is becoming more pronounced as industries recognize the potential of robots to perform complex tasks in various environments, highlighting the importance of integrating sophisticated sensory systems into robotic designs.
Features Science Sensors Technology AI infrastructure automation news
From audio tapes to AI: Interview with TDK investment director Ankur Saxena
Artificial intelligence has emerged as a leading focus in technology investment, yet some investors caution that the robotics sector may misinterpret the implications of recent advancements in large language models and generative AI. Ankur Saxena, the investment director at TDK Ventures, the corporate venture capital division of TDK, has voiced concerns regarding this trend. He emphasizes the need for a more nuanced understanding of how these AI breakthroughs can be effectively integrated into robotics, suggesting that a simplistic application of AI principles could lead to misguided strategies in the industry. Saxena's insights reflect a broader debate among investors about the future direction of robotics in light of AI developments, highlighting the importance of critical evaluation in investment decisions.
Features Financials & Investments Technology ai hardware ai robotics Ankur Saxena
China’s Alibaba unveils AI brains designed to power the next generation of robots
Alibaba, the Chinese technology giant, has unveiled its inaugural family of embodied AI models, marking a significant advancement in artificial intelligence technology. This launch, which took place recently, aims to enhance the interaction between humans and machines by integrating large language models with physical embodiments. The initiative is part of Alibaba's broader strategy to innovate and lead in the AI sector, responding to the growing demand for more intuitive and responsive AI systems. By leveraging its extensive data resources and expertise in machine learning, Alibaba seeks to revolutionize user experiences across various applications, from customer service to entertainment. The company plans to continue developing these models to further improve their capabilities and expand their use cases in the coming months.
AI and Robotics
Collecting robot training data is dirty, unglamorous work. Some AI labs are already paying XDOF to do it.
Recent discussions in the field of artificial intelligence highlight a significant challenge facing the development of physical AI systems. Experts emphasize that in order for physical AI to achieve milestones comparable to those of large language models (LLMs), a critical data issue must be addressed. As of October 2023, the existing datasets are insufficient to support the complex learning and operational needs of physical AI. This gap in data could hinder progress and innovation in creating AI that can effectively interact with and navigate the physical world. Addressing this problem is essential for advancing the capabilities of physical AI, ensuring that it can perform tasks with the same proficiency as its software counterparts.
AI Startups a16z robots Thrive Capital
Interview with Luo Jianlan: The true scaling law of robots occurs in real deployment loops.
In the past six months, the focus of the domestic embodied intelligence sector has shifted from hardware competition to the deeper challenges that define the intelligence limits of robots. Luo Jianlan, an associate professor at Shanghai Chuangzhi Academy and chief scientist at Zhiyuan Robotics, argues against the prevailing notion that robots can replicate large language models through sheer data accumulation. He emphasizes that the core issue in embodied intelligence is not about breakthroughs in isolated components but rather the ability to create a closed-loop system in real-world deployments. Luo, who has a background in both academia and industry, including roles at Google X and DeepMind, believes that many teams in the sector are not genuinely pre-training models but are instead engaged in mid-training or fine-tuning due to the scarcity of high-quality interaction data. He asserts that true embodied intelligence requires a scalable closed-loop system, where deployment leads to data collection, which in turn enhances model capabilities. His current focus includes developing scalable online post-training infrastructure, enabling robots to learn continuously in real-world environments, and creating a world model that predicts the consequences of actions rather than merely generating video. Luo suggests that the future of embodied intelligence hinges on successfully integrating these elements into a cohesive system, with significant advancements expected in the next 12 to 18 months. He believes that the first team to effectively implement a "deployment-data-iteration" cycle in semi-structured environments like convenience stores will gain a substantial competitive edge.
The Divergence of VLA in Autonomous Vehicles and Robotics: Insights from Xiaopeng and Jim Fan
A recent article examines the divergent trajectories of VLA (Vision-Language Agents) in the realms of autonomous driving and robotics, underscoring the distinct operational requirements inherent to each field. The analysis delves into the complexities of incorporating world models into VLA systems, revealing significant challenges that could impact the future development of artificial intelligence in these areas. The discussion emphasizes the necessity for specialized strategies that cater to the unique demands of autonomous driving and robotics, suggesting that a one-size-fits-all approach may not be viable for advancing AI technologies effectively.
Autonomous Driving Robotics AI World Models Machine Learning
Robots learn to anticipate chaos, but still fail to read a decidedly human signal
Researchers at Cornell University are exploring the integration of artificial intelligence into robotics to enhance social intelligence, enabling robots to interpret facial expressions, anticipate human needs, and interact effectively within societal contexts. In a recent study, the team evaluated vision language models (VLMs)—AI systems capable of processing and generating both visual and linguistic data. The research focused on assessing these models' ability to predict outcomes in tense scenarios depicted in short videos, such as a toddler precariously carrying an overflowing mug of coffee. This investigation aims to advance the development of robots that can better understand and respond to human emotions and behaviors, ultimately improving their functionality in everyday environments.
Robotics
"BioGeometry" secures hundreds of millions in strategic funding to create a "microscopic world model" in life sciences.
AI-native biotechnology company BaiAo Geometry has successfully secured several hundred million yuan in strategic financing, with investments led by the Shanghai Biomedical Innovation Transformation Fund, Guoke Investment, Dacheng Wisdom, and Xinglian Capital, alongside follow-on investments from GaoRong Capital and the Index AI Industry Innovation Fund. The funds will primarily support the ongoing development of their life sciences micro-world model, GeoFlow, and the advancement of their proprietary drug pipeline. Artificial intelligence is rapidly evolving along two main trajectories: digital AI, represented by large language and multimodal models, and physical AI, exemplified by autonomous vehicles and humanoid robots. Life AI is emerging as a promising frontier, a sentiment echoed by leading global investors and scientists. BaiAo Geometry's GeoFlow model, launched in 2024, aims to understand and design molecular interactions at an atomic level, enabling the creation of novel molecules that have never existed in nature. The company has iterated GeoFlow multiple times, achieving significant advancements in protein structure prediction and de novo design capabilities. By applying Test-Time Scaling technology, BaiAo Geometry enhances the success rate of protein designs without the need for extensive retraining. This innovation allows for the rapid generation and optimization of high-affinity binding molecules, significantly reducing the time and cost associated with traditional drug discovery processes. BaiAo Geometry has established over 20 business development collaborations with domestic and international pharmaceutical companies, focusing on high-specificity antibody design and vaccine development. The company is currently working on the next iteration of GeoFlow, which aims to expand modeling from individual molecules to entire molecular systems, further revolutionizing drug development in the biotechnology sector.
Interview with Workr Robotics CEO Ken Macken: ‘Paying for automation by the hour’
Industrial robotics is undergoing a significant transformation, driven by advancements in artificial intelligence, large language models, and embodied AI. This evolution has generated renewed interest in the development of robots capable of understanding, reasoning, and interacting with their physical environments. Notable partnerships, including collaborations between Google DeepMind and Boston Dynamics, have intensified discussions surrounding the potential for more sophisticated general-purpose robots. As these technologies continue to evolve, the industry anticipates a future where robots can perform a wider array of tasks, enhancing their utility across various sectors. The ongoing innovations suggest a promising trajectory for the integration of robotics into everyday life, potentially reshaping industries and improving operational efficiencies.
Features Industrial robots Robotics Software automation news automation roi
Alibaba's Qwen3.7-Plus supports text, video and imagery inputs at low cost of $0.4/$1.6 per 1M token — but it's proprietary
Alibaba has unveiled its latest AI large language model, Qwen3.7-Plus, this week, enhancing its Qwen family with advanced multimodal capabilities and a cost reduction of 60% compared to the previous text-only model, Qwen3.7-Max. Unlike earlier versions, Qwen3.7-Plus is available solely under a closed commercial license through proprietary APIs and Qwen Chat, marking a significant shift from Alibaba's previous strategy of offering open-source models. This change may disappoint users and enterprises, including major U.S. companies like Airbnb, that relied on open-source versions. The new model excels in multimodal tasks, such as generating enterprise-grade visuals and analyzing videos and images, which its predecessor could not perform. With a competitive pricing structure, Qwen3.7-Plus is positioned just above its Chinese competitor's discounted model. It features a 1-million token context window and a unique parameter called 'preserve_thinking,' which helps maintain continuity during complex tasks, a critical need for developers. Despite its advantages, benchmarks indicate that Qwen3.7-Plus still falls short of leading U.S. models in raw capability. However, it is designed to replace high-cost models in developer workflows and robotic process automation, offering a cost-effective solution for enterprises. The model's cloud-based deployment raises compliance concerns for organizations with strict data residency requirements, as it cannot be downloaded or hosted locally. Overall, Qwen3.7-Plus presents a compelling option for enterprises seeking efficient, multimodal AI solutions without incurring high operational costs.
Technology
Alibaba's Qwen Team Enters Embodied AI With Qwen-VLA Model
Alibaba's Tongyi Qianwen team has unveiled Qwen-VLA, marking the company's inaugural foray into the realm of vision-language-action models for embodied artificial intelligence. This launch, which took place recently, positions Alibaba to compete in the rapidly evolving sector of physical world AI. The development of Qwen-VLA is driven by the increasing demand for advanced AI systems capable of understanding and interacting with the physical environment, highlighting the company's commitment to innovation in artificial intelligence technologies. Through this new model, Alibaba aims to enhance the capabilities of AI in real-world applications, paving the way for more sophisticated interactions between machines and their surroundings.
AI
New Server Hopes to Break Through AI’s “Memory Wall”
Majestic Labs, an AI hardware startup, is addressing the memory limitations of large language models (LLMs) with its upcoming server, Prometheus, set to launch in 2027. This innovative server will feature up to 128 terabytes of memory, significantly surpassing the capabilities of Nvidia’s current offerings. Co-founder Sha Rabii emphasizes that this substantial memory increase will enhance performance and efficiency, particularly as models grow larger. Prometheus employs a unique DRAM-centric architecture, utilizing LPDDR6 memory and a proprietary memory interface with miniature copper cables that allow for greater memory placement flexibility. This design aims to overcome the “memory wall” that hampers LLM performance, providing a memory bandwidth of up to 25.6 terabytes per second. To complement its memory capabilities, Prometheus will incorporate the Ignite AI processing unit, which combines ARM application cores with RISC-V vector and tensor cores on a single chip. This integration allows for seamless handling of LLM inference tasks without the need for processor handoffs. Majestic Labs is also focused on ensuring compatibility with existing AI frameworks like PyTorch and OpenAI’s Triton, allowing customers to run their models without modifications. The server, designed in compliance with the Open Compute Project, will be modular, enabling future memory upgrades. Despite the advanced technology, Majestic Labs aims to offer competitive pricing by leveraging DRAM instead of more expensive high-bandwidth memory. Rabii claims that this approach could reduce customer capital expenditures and power consumption significantly, potentially by 10 to 50 times, depending on the workload.
Memory Server Ai-accelerators Performance
SAP SE (SAP), Cyberwave Deploy Autonomous AI Robots in Logistics Warehouse
SAP SE and Cyberwave have successfully deployed fully autonomous, AI-powered robots in SAP's logistics warehouse located in St. Leon-Rot, Germany, as of May 11, 2026. This initiative represents a significant advancement for SAP, transitioning its Physical AI technology from research to practical application. The robots, powered by SAP’s cloud-native Logistics Management solution and the SAP Business Technology Platform, are now capable of performing various tasks including box folding, packaging, and shipping fulfillment. The deployment addresses common challenges in logistics robotics, such as unpredictable environments and diverse object shapes that often hinder traditional systems. Cyberwave's innovative platform utilizes Vision-Language-Action and Reinforcement Learning models, enabling non-expert operators to teach robots new tasks through simple demonstrations. This approach significantly reduces training time from weeks to hours and allows robots to adapt to dynamic conditions in real-time. As a result of this integration, SAP has reported increased warehouse throughput and a decrease in physically demanding tasks for human workers. The project serves as a successful reference implementation, showcasing how a robust digital infrastructure combined with adaptive AI can enhance logistics operations. Both SAP and Cyberwave are now focused on further developing these Embodied AI capabilities to support future large-scale deployments.
NVIDIA Launches Alpamayo 2 Super Open Reasoning Model for Robotaxis
NVIDIA has unveiled its latest advancement in artificial intelligence, the Alpamayo 2 Super, a sophisticated vision language action model featuring 32 billion parameters. This new model is part of the expanding Alpamayo family, which includes a range of open AI models, simulation frameworks, and physical AI datasets designed to enhance safety in level 4 autonomous systems. The announcement was made today, showcasing NVIDIA's commitment to pushing the boundaries of AI technology. The Alpamayo 2 Super aims to improve the capabilities of AI in understanding and interacting with complex visual and linguistic inputs, thereby facilitating more advanced applications in various industries. This development reflects NVIDIA's ongoing efforts to lead in the AI space by providing robust tools for researchers and developers.
China’s new LLM-powered ‘AI brain’ automates satellite surveillance
Chinese aerospace researchers have unveiled an innovative system that utilizes Large Language Models (LLMs) to enhance various aspects of aerospace engineering. This development was announced during a recent conference focused on advancements in aerospace technology, held in Beijing. The researchers aim to improve design processes, streamline communication, and facilitate problem-solving in the aerospace sector through the application of artificial intelligence. The motivation behind this initiative stems from the increasing complexity of aerospace projects, which demand efficient and effective solutions. By integrating LLMs, the researchers hope to harness the power of AI to analyze vast amounts of data and generate insights that can lead to more innovative designs and improved operational efficiency. The system operates by processing extensive datasets related to aerospace engineering, enabling it to assist engineers in generating design concepts, optimizing workflows, and predicting potential challenges. This approach not only aims to reduce the time and resources required for development but also seeks to foster collaboration among engineers by providing a common platform for communication. As the aerospace industry continues to evolve, the introduction of such advanced technologies is expected to play a crucial role in shaping the future of aerospace engineering, making it more adaptive and responsive to the challenges ahead.
ugo and FastLabel launch training program for developing VLA model with domestic humanoids and physical AI.
Ugo Corporation and FastLabel Inc. have launched a hands-on training program aimed at facilitating the development of Vision-Language-Action (VLA) models for companies, universities, and research institutions. This initiative utilizes the domestically produced humanoid robot, the "ugo Pro R&D model," to support participants from the initial stages of model development. The program, titled "ugo VLA Model Development Training Program powered by FastLabel," is designed to enhance practical skills and knowledge in the emerging field of VLA technology.
Is Tesla, Inc. (TSLA) One of the Best ARK Stocks to Buy Right Now?
Tesla, Inc. has recently made headlines by increasing the prices of its Model Y lineup for the first time in two years. As reported by Reuters on May 16, 2026, the company raised the prices of its premium all-wheel and rear-wheel drive variants by $1,000, bringing them to $49,990 and $45,990, respectively. Additionally, the Model Y Performance All-Wheel Drive saw a $500 increase, now priced at $57,990. This adjustment follows a previous price hike in 2024, when all Model Y prices were raised by $1,000. Notably, Tesla had also increased the price of its high-end Cybertruck by $15,000 last August, despite facing challenges such as weak sales and recalls. In another development, former Tesla AI executive and OpenAI co-founder Andrej Karpathy announced on May 19 that he has joined Anthropic, expressing enthusiasm for contributing to the next phase of large language models. Tesla continues to operate in the electric vehicle and energy sectors, focusing on automotive and energy generation and storage technologies. While some analysts see potential in Tesla as an investment, they suggest that certain AI stocks may offer greater upside with less risk.
China’s AI start-up funding triples to US$16b in first quarter amid bets on LLMs, robotics
In the first quarter of the year, funding for artificial intelligence start-ups in China experienced a remarkable surge, increasing nearly threefold compared to the same period last year. Investors directed over 110 billion yuan (approximately US$16.2 billion) into these ventures, marking a 185 percent rise. This significant influx of capital is largely attributed to heightened enthusiasm surrounding large language models (LLMs) and embodied AI technologies, reflecting a growing confidence in the country's technology sector. The data, released by a Beijing-based research firm, underscores the accelerating interest and investment in AI as a key driver of innovation in China’s evolving tech landscape.
Robotics needs a service framework.
RSF defines a common language for robot service capability, lifecycle operations, certification pathways, and service-provider networks.
