Latest news and articles about research
While headlines chase humanoid robots and AI chatbots, a quieter revolution is happening on factory floors and construction sites. From MIT's carbon-cutting voxel assembly to ABB's turnkey surface finishing cells, the future of making things isn't about replacing humans—it's about redesigning how we build from the ground up.
Tutor Intelligence has built DF1, a Data Factory featuring 100 bimanual robotic manipulators trained by remote human tutors to develop its Ti0 vision-language-action model for real-world robot learning. The MIT-CSAIL-founded startup raised $34 million in Series A funding in December 2025 and claims DF1 is the largest robotic data factory in the U.S., using proprioceptive teleoperation and velocity normalization techniques to train generally capable robot AI at scale.
While enterprise software races to deploy AI agents for finance teams and office workflows, a quieter revolution is happening on construction sites and factory floors. Physical AI isn't just automating tasks—it's fundamentally redesigning how we build things.
Microsoft is presenting 11 papers at NSDI 2026, a leading conference on networked systems design and implementation, covering advances in AI systems, cloud infrastructure, and large-scale networks. Key contributions include DroidSpeak for efficient LLM inference through KV cache sharing, Eywa for automated protocol testing using LLMs, and AVA for video analytics with vision-language models.
Hiroshi Fujiwara and Robert Little have been named recipients of the 2026 Joseph F. Engelberger Robotics Awards for their contributions to robotics leadership and technology. Fujiwara, executive director of the Japan Robot Association, is recognized for strategic policy work and international collaboration, while Little, co-founder of ATI Industrial Automation, is honored for advancing robotic end-effector technology and practical automation adoption in manufacturing.
Beverly Da Costa is the inaugural graduate of Carnegie Mellon University's new Bachelor of Science in Robotics program, launched in 2023 by the Robotics Institute. The program represents a maturation of the robotics field, creating the first undergraduate degree focused on robotics alongside CMU's other computer science offerings, addressing growing industry demand for workers with diverse experience levels.
MIT researchers have developed a voxel-based robotic assembly system for construction that could reduce embodied carbon by up to 82% compared to traditional methods like 3D concrete printing and steel framing. The team created three new voxel designs and robotic assemblers called MILAbots that can efficiently place modular building blocks with snap-fit connections, with results published in Automation in Construction.
Amazon scientists and policy experts outline the company's responsible AI (RAI) pipeline, which integrates safety, fairness, and accountability throughout the AI development lifecycle across four phases: pretraining, post-training, evaluation, and frontier-risk assessment. The approach is supported by over 70 internal and external RAI tools, 500+ research papers, and tens of thousands of hours of employee training, guided by eight core pillars including safety, fairness, privacy, and transparency.
The Wall Street Journal reported that the offer could come as soon as this month.
Companies are paying everyday people to film mundane tasks and control robotic arms—a grassroots approach to training humanoid robots that sounds democratic but reveals a fundamental misunderstanding of how machines actually learn. The problem isn't quantity of data. It's that we're collecting movement without meaning.
Researchers at Science Tokyo have developed a ferroelectric tunnel junction memory chip using hafnium oxide that improves performance as it scales down to 25 nanometers, contrary to conventional miniaturization principles. By redesigning the device structure to reduce crystal boundary leakage and employing a new electrode fabrication method, the team created a memory device that consumes significantly less energy, with potential applications in ultra-efficient AI systems, wearables, and smartphones.
Astronomers at the University of Warwick have developed an AI system called RAVEN that discovered over 100 exoplanets in NASA's TESS mission data, including 31 newly confirmed worlds and 2,000+ high-quality candidates. The machine learning pipeline excels at distinguishing genuine planetary signals from false positives like eclipsing binary stars, enabling the identification of rare planet types including ultra-short-period worlds and those in the 'Neptunian desert.'
This article presents a phase stability regulator for autonomous mobile robots (AMRs) designed to manage computational instability in dynamic environments. The regulator uses two dynamic parameters—external task gradient (ΔN) and internal behavioral divergence (ΔD)—to prevent combinatorial overload and maintain real-time behavioral stability without sacrificing planning capability.
Three recent developments in robot navigation reveal a field pulling in opposite directions — bio-inspired algorithms from insects, neural networks trained on human language, and bat-like ultrasound systems. The lack of consensus might actually be progress.
While self-driving cars fight regulatory battles on land, underwater robots are quietly revolutionizing ocean exploration and commerce. From deep-sea mining to whale research, autonomous underwater vehicles are proving that the real frontier for mobile robotics might be beneath the waves.
Carnegie Mellon University's Robotics Institute is launching the next phase of the Vision-Language-Navigation (VLN) Challenge, which tasks teams with developing robots that can understand natural language instructions and navigate complex 3D environments without relying on predefined ground truth data. This year's challenge emphasizes semantic and spatial reasoning, requiring robots to interpret raw sensor data and understand object relationships in real-world settings, with competition beginning in simulation before advancing to physical robot testing.
Robot Talk Episode 154 features an interview with Andrew Philippides, Professor of Biorobotics at the University of Sussex, discussing how visual navigation systems in insects like ants and bees can inspire improved robot navigation algorithms. Philippides combines biological experiments, robotics, modeling, and machine learning to develop novel AI and biorobotic algorithms based on understanding natural navigation and learning abilities.
The article examines how deformable materials like fabric represent a critical manufacturing challenge for physical AI systems, arguing that successful automation requires redesigning garment assembly processes rather than simply replicating traditional sewing workflows. It discusses how robots struggle with fabric's unpredictable behavior (stretching, wrinkling, deformation) and emphasizes that solving this problem requires integrated design of robotics, perception systems, and control methods working together in production environments.
Orpheus Ocean, a spinoff from Woods Hole Oceanographic Institution, is deploying low-cost autonomous underwater vehicles (AUVs) designed for deep-sea exploration and sediment sampling at depths up to 11,000 meters. The submersibles, built for a couple hundred thousand dollars compared to $5-10 million for existing alternatives, are conducting their largest operational test by mapping the Pacific seafloor for critical mineral deposits while collecting sediment cores and high-resolution imagery.
This past week delivered another gut punch for science in the US. This time, the target was the National Science Foundation-a federal agency that funds major research projects to the tune of around $9...
Researchers conducted red-team testing of a network of over 100 AI agents to identify vulnerabilities that emerge only through agent-to-agent interaction at scale, discovering four network-level risks: propagation (agent worms spreading and collecting data), amplification (exploiting trusted agents to spread false claims), trust capture (hijacking verification systems), and invisibility (obscuring attack sources). The study reveals that single-agent benchmarks fail to predict network behavior, and while some agents demonstrated defensive behaviors, securing multi-agent systems remains an open challenge as AI agents increasingly interact across platforms like email and GitHub.
While headlines trumpet humanoid deployments and AI breakthroughs, the unglamorous world of tactile sensing might be what actually unlocks industrial automation at scale. Companies are quietly solving the touch problem that's kept robots clumsy for decades.
This article discusses the development of reliable AI tools for healthcare, specifically focusing on AI co-clinician systems that could enable new models for clinical care. The research appears to be from July 2023 and addresses the challenges of creating trustworthy AI applications in medical settings.
A new study from Zhejiang University challenges claims that the AI model Centaur can replicate human cognitive behavior, revealing it may be memorizing patterns rather than understanding concepts. Researchers demonstrated that Centaur fails at true language comprehension when tested with modified prompts, suggesting major limitations in current large language models' ability to achieve genuine understanding.
XELA Robotics announced two enhancements to its uSkin tactile sensor family: magnetic interference compensation for handling ferromagnetic materials in industrial settings, and CAN FD protocol support enabling higher-speed communication and more sensors per gripper. The company will showcase these advancements at the Robotics Summit & Expo in Boston in May 2026, with deliveries expected to begin in Q3 2025 and May 2026 respectively.
Amazon researchers demonstrate three attacks that can extract private training data from AI models—membership inference, gradient-based reconstruction in federated learning, and data extraction from shared models—and show how differential privacy and secure multiparty computation provide effective cryptographic defenses against these threats.
While the tech world obsesses over chatbots and image generators, a quieter breakthrough is reshaping robotics and AI: world models that let machines actually understand physics, not just text. From Berkeley labs to YouTube's AI features, the shift from language to physical understanding is accelerating faster than most realize.
Software engineer Chris Nager created a playable Doom application that runs within AI chatbots like ChatGPT and Claude by leveraging Model Context Protocol (MCP), an open-source standard that allows LLMs to connect with external tools and data sources. Users can launch the game by typing "play Doom" directly in the chatbot interface, with fallback URL access if inline launching isn't supported by the client.
Researchers at Worcester Polytechnic Institute developed an ultrasound-based perception system for tiny drones inspired by bat echolocation, enabling navigation in low-visibility environments where cameras and lidar fail. The system combines an acoustic shield that reduces propeller noise with a neural network called Saranga that recovers weak echo signals, allowing small aerial robots to navigate safely in smoke, dust, darkness, and other challenging conditions for search-and-rescue operations.
Project CETI has deployed autonomous underwater gliders developed by Alseamar to track and listen to sperm whale vocalizations, integrating acoustic monitoring algorithms to follow whales beneath the ocean surface. The gliders use passive buoyancy-based locomotion to operate quietly for extended periods while collecting acoustic data, overcoming limitations of previous monitoring methods like buoys, boats, and short-term tags.
Google's classified AI deal with the Department of Defense includes unusual restrictions on autonomous weapons and mass surveillance—revealing that the government feels it needs explicit guardrails against its own contractor. This isn't partnership. It's containment.
NVIDIA unveiled Nemotron 3 Nano Omni, an open multimodal AI model that unifies vision, audio, and language processing in a single system for more efficient AI agents. The 30B parameter model achieves 9x higher throughput than competing open omni models while reducing latency, cost, and context fragmentation, with applications in computer use agents, document intelligence, and audio-video reasoning workflows.
NVIDIA Omniverse and OpenUSD are enabling manufacturers to shift from real-world testing to high-fidelity simulation for training AI and robotic systems, with companies like ABB Robotics achieving 99% sim-to-real accuracy and reducing commissioning times by up to 80%. The article discusses SimReady as a content standard for physically accurate 3D assets, and showcases practical applications including aerodynamic simulation compression and real-time factory intelligence systems.
Researchers from BAIR introduce GRASP, a gradient-based planning algorithm designed to make long-horizon planning with learned world models more practical and robust. The method addresses key challenges in using large-scale world models for control by lifting trajectories into virtual states for parallel optimization, incorporating stochasticity for exploration, and reshaping gradients to improve signal quality in high-dimensional vision spaces.
While academic institutions rush to partner with tech giants on cutting-edge AI research, their own digital infrastructure is crumbling. Recent security breaches at top universities reveal a troubling paradox: we're training tomorrow's robots on yesterday's systems.
The 2026 Robotics Summit & Expo (May 27-28 in Boston) will feature an extensive track dedicated to Physical AI, covering topics including scalable robot systems, data integration, embodied AI, large behavior models, and industrial applications. The event will include keynotes and panel discussions from leading researchers and industry figures such as Russ Tedrake, executives from companies like Brain Corp, Agtonomy, and Universal Robots, and demonstrations of AI-powered robotics integration across retail, logistics, and manufacturing sectors.
Amazon researchers and UIUC colleagues introduced the C3LLM framework, a statistical method for assessing catastrophic failure risks in large language models during multi-turn conversations. The framework uses graph-based modeling and Clopper-Pearson confidence intervals to compute probabilistic bounds on attack success rates, addressing limitations of traditional red-teaming approaches that focus on isolated prompts rather than conversational contexts.
Hospitals are rapidly deploying AI diagnostic tools, scribes, and predictive systems—but researchers are raising a troubling question: does any of it actually help patients? The gap between technical accuracy and real-world outcomes reveals a healthcare industry racing ahead of its own evidence.
DeepSeek released its V4 flagship AI model with improved long-context processing capabilities that match leading closed-source competitors, while being optimized for Huawei's Ascend chips. The article also highlights the growing importance of 'world models' in AI research—systems that understand physical environments rather than just digital ones—which researchers like Fei-Fei Li and Yann LeCun argue are crucial for advancing AI applications in robotics and real-world tasks.
OpenAI has released Symphony, an open-source specification for orchestrating AI coding agents that uses project management tools like Linear as a control plane. The system allows multiple coding agents to autonomously pull tasks from an issue tracker and work continuously with human review, achieving a 500% increase in pull request throughput by eliminating the bottleneck of managing interactive coding sessions.
Researchers at the University of Chicago developed penPal, a robotic drawing assistant that uses tendons and servos to execute programmed pen movements while artists control larger canvas movements. A study with ten participants demonstrated that artists discovered novel interaction methods with the tool, with some describing the experience as collaborative drawing with another person.
SS Innovations' drone-deployed surgical platform represents a pivotal shift in battlefield medicine—and raises uncomfortable questions about how automation changes the calculus of conflict.
The Robot Report Podcast episode features Dr. Jan Liphardt, a Stanford bioengineering professor and founder of OpenMind and OM1, discussing the future of physical AI. The episode covers Tesla's announcement of a large-scale Optimus humanoid factory in Fremont with plans for a second-generation facility in Texas targeting 10 million units annually, a patent infringement case involving Teradyne Robotics and Elite Robots over collaborative robot software, and HII's partnership with Path Robotics and GrayMatter Robotics on a high-yield production robotics program.
Hundreds of subdomains from dozens of universities have been hijacked by scammers.
Sony's table tennis robot just defeated elite human players in competitive matches — a historic first in physical sports. But the real story isn't about winning. It's about what happens when we finally stop treating human-machine competition as the endgame.
Robot Talk Episode 153 features an interview with Chenying Liu from the University of Oxford about origami-inspired robots and embodied physical intelligence. Liu's research explores how robot design can integrate geometry, materials, and control to enhance autonomy and robustness by embedding intelligence directly into physical structures.
Researchers at the University of Michigan and University of Toronto argue in Nature Medicine that while AI tools are rapidly being deployed in healthcare settings—including AI scribes for documentation, predictive tools, and diagnostic aids—there is insufficient evidence that these technologies actually improve patient health outcomes. The paper highlights a critical gap: healthcare providers are widely adopting AI without rigorously assessing clinical impact, bias, or unintended consequences on medical decision-making and patient care.
Physicists at Emory University used a custom neural network to discover new physics in dusty plasma systems, achieving over 99% accuracy in describing non-reciprocal forces and revealing previously hidden patterns that challenge long-standing theoretical assumptions. The AI approach demonstrates that machine learning can uncover entirely new physical laws in complex many-body systems, with potential applications extending to industrial materials and biological systems.
Fusion power could provide a steady, zero-emissions source of electricity in the future-if companies can get plants built and running. But a new study suggests that even if that future arrives, it mig...
Researchers at the University of Cambridge have developed a brain-inspired nanoelectronic device using modified hafnium oxide that could reduce AI energy consumption by up to 70%. Unlike conventional chips that shuttle data between memory and processing units, this memristor-based device combines both functions in one location, mimicking how neurons work and operating at switching currents a million times lower than conventional oxide-based memristors.
