Mathematicians Warn AI Could Erode Their Field: Sixteen mathematicians have issued the Leiden Declaration on Artificial Intelligence and Mathematics, warning that unchecked AI could reshape their discipline in damaging ways. The 11-page statement, endorsed by the International Mathematical Union, argues that mathematics is not simply about producing correct answers but also about creativity, understanding, collaboration, and knowledge pursued for its own sake. The authors worry that commercial AI incentives conflict with those values. They warn that AI-generated papers could flood peer review, muddy credit for discoveries, and disadvantage researchers who avoid AI tools. They also raise concerns that mathematical work may be used to train military or surveillance systems. The declaration is open for signatures and will be discussed at next month’s International Congress of Mathematicians in Philadelphia. (Science)
Meta’s AI Overhaul Shows Early Signs of Progress: A year after Mark Zuckerberg recruited Scale AI co-founder Alexandr Wang to revive Meta’s struggling artificial intelligence program, the company has produced Muse Spark, its most credible model yet. Wang, then 28, was given unusual autonomy to build TBD Lab, a secretive team of about 100 elite researchers working inside Meta’s Menlo Park headquarters. Supporters say the group moved quickly, delivering a model that shows strength in visual understanding and could improve Meta’s ads, assistants, avatars, wearables, and business tools. Critics argue the progress is incremental, note tensions with existing AI teams, and say Muse Spark still trails rivals in coding. Future Meta models are expected to focus on coding, agentic tasks, multimodal abilities, and video generation. (Ars Technica)
Nvidia Pushes Humanoid Robot Research Beyond China: Nvidia is expanding its humanoid robotics strategy beyond its work with China’s Unitree, with plans to collaborate with robot makers in the United States, Europe, and South Korea. The company is helping build standardized research robots that combine Unitree’s H2 body, Sharpa’s robotic hands, and Nvidia computing hardware. Stanford University and UC San Diego are among the institutions expected to use the systems. Nvidia says the project is partly about cybersecurity: software updates would pass through Nvidia chips, allowing code authentication and protections similar to those used in data centers. The move comes as U.S. lawmakers scrutinize Unitree’s alleged ties to China’s government and military, making international diversification strategically important. (Reuters)
LLMs Help Robots Understand What Humans Leave Unsaid: MIT CSAIL researchers have developed a system that uses large language models to help robots interpret vague human instructions and focus on the details that matter. The method, called Masked Inverse Reinforcement Learning, expands ambiguous commands based on a human demonstration, then identifies which objects or environmental features are relevant to the task. In tests, robots learned manipulation tasks with nearly five times less demonstration data and inferred unstated user preferences up to 15% better than comparable systems. A robot could, for example, understand that “stay away” means avoiding a laptop while delivering a cup. The work will be presented at ICRA 2026 and points toward more natural human-robot teaching. (TechXplore)
Foundation Models Could Transform Robot Swarms: A new Science Robotics viewpoint argues that foundation models could fundamentally change how robot swarms are designed and deployed. Traditional swarm systems rely on manually programmed rules that can be brittle when robots encounter unexpected events, such as sensor failures, disaster-zone obstacles, or shifting mission priorities. Researchers from Université Libre de Bruxelles and CISPA Helmholtz Center for Information Security propose embedding foundation models directly into individual robots so they can process sensor inputs and generate collective actions. In principle, a forest-monitoring swarm could redirect itself toward rescuing an injured person without being explicitly programmed for that scenario. The authors also warn that hallucinations, hardware limits, and security risks must be addressed. (EurekAlert)
Virtual Tomato Farms Train Harvesting Robots: Osaka Metropolitan University researchers have built a method for creating realistic virtual tomato farms to train agricultural AI systems. Harvesting robots already use object detection to locate fruit and judge ripeness, but training those systems usually requires large amounts of labeled field data, which is slow and labor-intensive to collect. The new approach reconstructs a virtual farm from camera images gathered by agricultural robots, then automatically generates realistic tomato images and corresponding training labels. By simulating varied farm conditions, the system could help robots learn perception tasks before they enter real fields. The work addresses a major bottleneck in robotic agriculture: making harvesters reliable enough for messy, changing crop environments. (EurekAlert)
Four-Armed Space Robot Built for Microgravity: Orbit Robotics, an ETH Zurich spinout, has designed Helios, a four-armed robot intended to help astronauts in microgravity. Unlike humanoids modeled for walking on Earth, Helios is built around the demands of weightlessness. Two arms can anchor the robot to interior surfaces while the other two manipulate tools, cargo, or equipment. The design uses tendon-driven limbs, concentrating motors near the shoulders to reduce distal weight, and incorporates rolling-contact elbows for smooth motion. That matters in microgravity, where sudden jerks can destabilize both robot and payload. The company frames Helios as a way to shift repetitive station maintenance away from astronauts, freeing crew time for scientific work. (New Atlas)
Qualcomm Offers Production-Ready Robotics Platform: Qualcomm is introducing the Dragonwing IQ10 Robotics Reference Design, a robotics platform meant to help companies move from prototypes to deployable machines. Built around the Dragonwing IQ10 processor, the system combines compute, sensors, networking, software, and cooling in one package. It supports up to 12 GMSL2 cameras, LiDAR, time-of-flight sensors, IMUs, ROS2, and deterministic interfaces for motion control. Qualcomm says the platform can handle perception, planning, reasoning, navigation, manipulation, task orchestration, and natural-language interaction on-device. It is aimed at industrial robots, autonomous mobile robots, and humanoids, with global availability planned for September 2026. Early partners include NEURA Robotics, Advantech, Thundercomm, VinMotion, and others. (Times of India)
Robot Dogs Join World Cup Security Plans: Hyundai and Boston Dynamics are deploying two robot dogs to support World Cup security operations. One Spot robot will be stationed at the tournament’s International Broadcast Center in Dallas, while another will assist at MetLife Stadium in New Jersey, which is hosting eight matches. The quadruped robots use cameras to help monitor assets and identify possible security risks, reporting information back to human supervisors. Their deployment is part of a much larger security plan involving more than 30 law-enforcement agencies in North Texas, including both uniformed and undercover officers. The story captures the growing use of mobile robots in event security while underscoring that they remain supplements to human teams. (Axios)
Computex Shows Robotics Hype and Reality: Robots were a major theme at Computex 2026, where companies showcased humanoids that danced, flipped, pulled heavy objects, and drew crowds. But the demonstrations also revealed the gap between viral robotics and practical autonomy. Many of the most impressive robots were remotely operated or pre-programmed rather than fully autonomous. Nvidia announced a humanoid research partnership with Unitree, while Asus presented AI-powered service robotics aimed at care and support tasks for elderly users and others. The robotics push sat alongside broader Computex themes: agentic AI, supply-chain strain, booming AI chip demand, and geopolitical concerns around Taiwan’s semiconductor role. The takeaway: physical AI is advancing, but autonomy remains hard. (TechXplore)
Reusable Robot Skills Could Cut Programming Burden: A new International Journal of Robotics Research paper proposes a way to make robot manipulation software more reusable across different machines. The authors focus on Learning-from-Observation, in which robots learn tasks by watching human demonstrations. Their system separates the task representation from the specific robot hardware, then uses a standard library of “skill agents” to translate those representations into robot commands. The goal is to reduce the need to write custom control programs every time a task moves to a different robot body or gripper. The researchers demonstrate that the same task representations can work across Nextage and Fetch robots, as well as different end-effectors, including Shadow Hand-Lite and a parallel gripper. (Sage)
Soft Robots Get Plant-Inspired Tactile Sensing: Soft Robotics published new work on plant-inspired elastic-hydraulic tactile sensing for soft robots, highlighting continued progress in machines that can safely interact with complex objects, living systems, and human-built environments. The paper, first published May 26, focuses on quantitative stiffness estimation, a key capability for soft robotic systems that need to handle delicate or variable materials. Better stiffness sensing could help soft grippers decide how firmly to hold fruit, tissue, tools, or irregular objects without crushing them or losing control. The journal’s latest issue also includes research on a bioinspired swallowing soft gripper with optical waveguides, underscoring a broader trend: soft robots are gaining richer sensing, not just more compliant bodies. (Sage)