Latest news and articles about research
Sony's table tennis robot just beat world-class human players. This isn't about ping-pong — it's about robots finally mastering the kind of split-second, unpredictable physical interactions that have stumped them for decades.
Anduril and Meta are developing augmented-reality military headsets that use AI, eye-tracking, and voice commands to help soldiers coordinate drone strikes and tactical decisions. The project, called Soldier Born Mission Command (SBMC), represents a $159 million Army contract to create glasses that overlay tactical information and use large language models (Google Gemini, Meta Llama, Anthropic Claude) to translate soldier speech into actionable military commands, with prototypes expected years away from deployment.
Sony AI's table tennis robot Ace defeated some of the world's best players in recent evaluations, winning 3 out of 5 matches against elite competitive athletes, marking a significant advancement in AI systems operating in fast, uncertain real-world environments. The robot uses event-based vision sensors, high-speed cameras, and deep reinforcement learning to handle complex challenges like spin estimation and real-time ball tracking, demonstrating progress beyond controlled digital environments where AI has traditionally excelled.
Researchers at Tokyo University of Science have developed an AI-powered physics model called the entropy-feature-eXtended Ginzburg-Landau (eX-GL) model to understand and reduce energy loss in electric motors caused by complex magnetic domain structures. The model uses persistent homology and machine learning to analyze the chaotic 'maze-like' magnetic patterns in soft magnetic materials, revealing how temperature and microscopic structures trigger energy waste as heat, with potential applications for improving electric vehicle efficiency.
The release of Asimov v1 at $15,000 marks a turning point for humanoid robotics — not because it's cheap, but because it's hackable. When thousands of developers can finally afford to experiment with bipedal platforms, the question isn't what companies will build, but what chaos independent builders will unleash.
Asimov v1, an open-source humanoid robot kit, is now available with a target price of $15,000 USD, significantly lowering the barrier to entry for humanoid robotics development compared to corporate projects. The robot features 25 degrees of freedom, runs on Raspberry Pi 5 and Radaxa CM5 compute, and includes open-source hardware designs on GitHub to enable hobbyists and researchers to build and customize their own humanoid systems.
A roundup of recent reviews published by Engadget.
Microsoft researchers present findings from their paper on AI delegation reliability, examining how language models preserve semantic content across long-horizon delegated workflows. The study finds that frontier models experience 19-34% degradation in artifact fidelity over 20 delegated iterations, though Python workflows show greater robustness, highlighting reliable long-horizon delegation as an important open research challenge.
Researchers present a new scaling law framework that relates architectural design choices in large language models to inference efficiency, demonstrating how to improve throughput by up to 47% without sacrificing accuracy. The work, presented at ICLR, identifies optimal configurations such as the MLP-to-attention ratio for models like LLaMA-3.2, addressing a gap in existing scaling laws that don't account for internal architectural decisions affecting inference performance.
Robot Talk Episode 156 features an interview with Gavin Kenneally, Co-Founder and CEO of Ghost Robotics, discussing rugged legged robots designed for dangerous missions in defence, security, and public safety applications. The episode explores how advanced robotics systems like Ghost Robotics' Vision 60 Q-UGV are being developed for challenging real-world environments.
Security researchers used Anthropic's Claude Mythos Preview AI system to identify vulnerabilities and develop a privilege escalation exploit for macOS on M5 silicon chips. The work demonstrates how advanced AI systems can help discover previously unknown security bugs and attack paths, with Mythos part of Anthropic's Project Glasswing initiative aimed at using AI to prevent cyberattacks.
NASA is developing a radiation-hardened AI space chip that delivers up to 100 times more computing power than current spacecraft processors while surviving deep space conditions. The High Performance Spaceflight Computing processor, currently undergoing rigorous testing at NASA's Jet Propulsion Laboratory, could enable spacecraft to operate more autonomously and support faster scientific analysis for future missions to the Moon and Mars.
A maker successfully demonstrates directional control of a vibrobot using only a single vibrating motor by rapidly switching the motor's rotation direction to approximate straight-line movement. The design, based on academic research, uses an ESP-01F microcontroller and Wi-Fi remote control to switch motor direction every 250 milliseconds, enabling the robot to follow traced paths despite slightly erratic movement.
From OpenAI's secure coding sandboxes to automated prompt engineering, the latest wave of AI development tools promises to write better code faster. But as these systems become more autonomous and opaque, we're approaching a threshold where even experienced developers can't fully explain what their tools are doing—or guarantee what they'll do next.
Promptimus is an automated prompt-engineering framework that optimizes already well-developed LLM prompts without manual intervention by identifying specific failure points and suggesting targeted improvements. The method achieves best results on 16 of 20 benchmarks while outperforming six leading automatic prompt optimization approaches, demonstrating model-agnostic generalizability and sample efficiency across various enterprise tasks.
When generative AI first moved from research labs into real-world business applications, enterprises made a tacit bargain: "Capability now, control later." Feed your proprietary data into third-party...
The Robotics Summit & Expo in Boston (May 27-28) will feature a session on humanoid robotics scaling, led by Texas Instruments' Giovanni Campanella, covering system-level innovations in sensing, processing, power management, and edge AI for next-generation humanoid robots. The summit will include 50+ technical sessions and 70+ speakers from leading robotics and AI companies, addressing design challenges like sensor fusion, real-time processing, and motor control critical to humanoid performance.
When Jennifer got a job doing research for a nonprofit in 2023, she ran her new professional headshot through a facial recognition program. She wanted to see if the tech would pull up the porn videos...
A hobbyist integrated a LiDAR matrix sensor (64-element array) into Zippy, a 3D-printed ESP32-based robot, to enable autonomous navigation by detecting obstacles and floor features in 3D space. The project demonstrates time-of-flight sensor capabilities for robotics perception, though the sensor mounting resulted in significant data loss and required optimization to process the 3D distance measurements effectively.
Microsoft introduces GridSFM, a lightweight foundation model that solves AC optimal power flow problems in milliseconds, enabling real-time optimization of electrical grids. The model uses a block-structured discrete neural operator to handle grids from 500 to 80,000 buses, potentially reducing up to $20B annually in congestion losses and unlocking 3.4 TWh of curtailed renewable energy.
This article discusses GMSL (Gigabit Multimedia Serial Link) diagnostics technology for autonomous and robotic vision systems, explaining how built-in diagnostic capabilities at the chip, link, and data layers enable early detection of sensor failures and signal degradation. The piece covers diagnostic requirements across the system lifecycle—from design and integration through production, runtime operation, and field service—to ensure robust and reliable perception systems in autonomous vehicles and robotics.
This article examines the evolution of vision connectivity technologies in robotics, tracing the progression from USB and Ethernet-based camera interfaces to Gigabit Multimedia Serial Link (GMSL). The piece discusses how GMSL, originally developed for automotive applications, offers advantages like deterministic low-latency transport, better scalability for multi-camera systems, and improved mechanical robustness compared to legacy interfaces, making it better suited for modern robotics applications requiring real-time perception and sensor fusion.
Researchers at Leiden University have developed microscopic flexible robots that move autonomously without sensors, software, or external control, instead relying on shape-environment interactions inspired by biological microswimmers. These 3D-printed chain-like microrobots demonstrate emergent intelligent behaviors including obstacle avoidance and object manipulation, with potential applications in targeted drug delivery and minimally invasive medical procedures.
This article examines GMSL (Gigabit Multimedia Serial Link) technology and its growing adoption in robotic vision systems for navigation and spatial awareness. GMSL, originally developed for automotive applications like ADAS and autonomous driving, is increasingly being deployed in production robots including warehouse AMRs, humanoid robots, and agricultural platforms to handle multi-camera synchronization, reduce cabling complexity, and provide deterministic latency—with roughly one-third of robotic opportunities now using or considering GMSL-based solutions.
This article examines the technical requirements for humanoid robots to safely interact with humans, focusing on vision and audio processing capabilities. It discusses how humanoid robots use sensors, processors, and technologies like GMSL for real-time visual data processing and intelligent hearing to maintain situational awareness and respond appropriately in shared human environments.
Researchers from Baycrest, University of Toronto, and York University discovered that AI can analyze natural speech patterns—including pauses, filler words, and word retrieval difficulties—to predict cognitive performance and detect early signs of dementia. The study found that speech timing characteristics are closely linked to executive function and could offer a practical, non-invasive alternative to traditional cognitive testing for monitoring brain health across the lifespan.
This article discusses how perception systems are converging across autonomous vehicles, robots, drones, and humanoids, with all platforms adopting distributed sensor networks and real-time processing architectures similar to automotive systems. The piece emphasizes that perception has evolved from a supporting feature to critical infrastructure in modern robotics, requiring integrated solutions for vision, audio, and tactile sensing with deterministic connectivity, diagnostics, and reliability to operate safely in dynamic, unstructured environments.
A security researcher conducted a detailed analysis of Unitree robot dogs, exposing multiple vulnerabilities including arbitrary command execution via Wi-Fi password fields, suspicious network traffic to Chinese servers, and sensory limitations that severely restrict practical applications. The investigation revealed that despite their affordable pricing and capable hardware, these quadruped robots have significant security flaws and firmware concerns that make them unreliable for real-world deployment.
MIT Technology Review highlights world models as an emerging area of AI gaining significant attention, with a focus on how AI systems can better understand and reason about the real world. The article features discussion from AI experts including Yann LeCun and covers the implications of world models for the future of AI development.
Google researchers introduce MatterSim-MT, a multi-task foundation model for materials science that accelerates computational materials design. The team reports experimental validation of MatterSim predictions by synthesizing tetragonal tantalum phosphorus with thermal conductivity matching silicon, plus 3-5x performance improvements in simulation speed integrated with LAMMPS for large-scale GPU-enabled computations.
Hello Robot announced Stretch 4, an upgraded version of its mobile manipulation platform now priced at $29,950, featuring twice the speed of Stretch 3, extended reach, and a comprehensive sensor array with dual 3D lidars and multiple cameras for improved safety and autonomy. The robot is designed for general-purpose physical AI applications in dynamic environments like homes, with enhanced sensing capabilities to detect obstacles and people while performing manipulation tasks.
Teams use Codex with GPT-5.5 to ship production systems and turn research ideas into runnable experiments.
Parameter Golf was an open machine learning challenge that attracted over 1,000 participants and 2,000+ submissions to optimize neural network models under strict constraints (16 MB artifact size, 10-minute training budget). The challenge revealed how AI coding agents are transforming machine learning research by lowering experimentation costs and accelerating competition pace, while also surfacing insights about model optimization techniques including training optimization, quantization, test-time training strategies, and novel modeling approaches.
Three separate research breakthroughs this week point to the same conclusion: the robotics industry is abandoning rigid programming in favor of adaptive learning systems. From dexterous hands to cross-platform skill transfer, the message is clear — tomorrow's robots need to figure things out on their own.
RLWRLD unveiled RLDX-1, a dexterity-first foundation model designed specifically for high-degree-of-freedom robotic hands in industrial applications. The model integrates vision, force sensing, temporal reasoning, and memory to handle complex manipulation tasks like pouring, object tracking, and precise contact control across single-arm, dual-arm, and humanoid embodiments.
SocialReasoning-Bench is a new benchmark that evaluates whether AI agents can effectively negotiate and act in users' best interests in social contexts like calendar coordination and marketplace negotiations. The research finds that current frontier models often complete tasks but frequently accept suboptimal outcomes, failing to advocate effectively for users even with explicit guidance, highlighting a gap in social reasoning capabilities that mirrors principal-agent relationships in law and economics.
Researchers used a Unitree G1 humanoid robot to provide haptic feedback in a VR driving simulator by having the robot grasp and move a player's chair in real-time based on G-force signals from the game. The study found participants enjoyed the synchronized motion feedback, though some reported fatigue and VR sickness from sustained vibrations, demonstrating the potential of humanoid robots as flexible, general-purpose devices for interactive applications.
Researchers at EPFL have developed Kinematic Intelligence, a novel framework that enables robots with different mechanical structures to learn and execute the same tasks from human demonstrations. The method converts human-demonstrated skills into general movement strategies and automatically adapts them to different robot designs while ensuring safe operation, demonstrated successfully with three different commercial robots performing an assembly line task.
Despite years of digitization, organizations capture less than one-third of the value expected from digital investments, according to McKinsey research. That's because most big companies begin with te...
Researchers at Jülich Supercomputing Centre and NVIDIA achieved a record-breaking milestone by simulating a 50-qubit quantum computer using JUPITER, Europe's first exascale supercomputer, surpassing the previous 48-qubit record. The breakthrough demonstrates how advanced supercomputing enables quantum research validation and algorithm development, with the simulation requiring 2 petabytes of memory and managing over 2 quadrillion complex numerical values across thousands of computing nodes.
A webinar announcement focusing on engineering robot tracks and 7th-axis linear motion systems for harsh, contaminated industrial environments such as welding cells, paint booths, and washdown areas. The session covers failure modes, design strategies, and preventive maintenance approaches to improve reliability and uptime in demanding real-world robotics applications.
The space agency is planning to put astronauts on the moon in 2028.
Researchers at MIT, led by Miana Smith, have developed MILAbot, an open-source inchworm robot capable of assembling building structures from modular voxel blocks without a traditional fixed base. The robot uses a unique locomotion approach where it anchors to completed structure sections while placing new blocks, offering potential advantages in embodied energy and labor costs compared to conventional construction methods, though current balloon-frame building remains more cost-effective.
Eric Chan, Chief Scientist at Rhoda AI, discusses how Direct Video Action (DVA) models trained on internet video data are replacing traditional robotics data collection methods. The approach enables robots to learn complex tasks with minimal training data and supports zero-shot learning capabilities for real-world deployment.
From OpenAI's Codex to Google's AlphaEvolve, AI systems are increasingly writing, optimizing, and deploying code autonomously. While the productivity gains are undeniable, the automation of software development raises urgent questions about security, accountability, and what happens when the code breaks.
Microsoft Research is excited to release an open dataset of approximate transmission topology of the U.S. power grid derived from publicly available data. The ability to study transmission-level power...
Carnegie Mellon University researchers developed World2Rules, an AI system that learns interpretable safety rules from airport operational and incident data to predict and prevent aircraft collisions. The system analyzes patterns of unsafe aircraft behavior using two years of FAA data from 42 U.S. airports and explains collision risks in understandable terms to enhance air traffic safety.
MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what's coming next. You can read more from the series here. Eight passengers aboa...
QNX, a BlackBerry division, will demonstrate its real-time operating system (RTOS) for robotics at the Robotics Summit & Expo in Boston on May 27-28, showcasing how its software enables safe, deterministic AI-driven control for industrial and collaborative robots. The company will present live demonstrations of robotic arm prototyping, digital factory automation with sensor fusion, and AI-based motion replication, while also launching a new benchmark report surveying 1,000 robotics developers on industry trends and challenges.
MIT Technology Review's newsletter covers multiple AI and robotics developments: an essay on 'AI malaise' exploring societal uncertainties around widespread AI adoption; advances in robot learning through AI that enable machines to adapt to real-world environments rather than relying on rigid programming; and applications of AI and robots in IVF technology. The newsletter also includes news briefs on AI economic impacts, ICE's smart glasses facial recognition plans, and China's competitive AI models.
