Mind Captioning Turns Brain Scans into Sentences

A new “mind captioning” technique brings noninvasive mind-reading closer to practice by translating fMRI brain activity into detailed natural-language descriptions of what a person sees or recalls. Researchers first mapped captions from 2,000+ videos into numerical “meaning signatures” using a large language model. They then trained a decoder to match participants’ brain-activity patterns to those signatures while watching clips, and used a separate text generator to output the closest sentence. The system progressed from vague to precise phrases—e.g., eventually describing “a person jumps over a deep waterfall”—and also worked when participants remembered scenes, implying similar neural representations for viewing and memory. Potential applications include communication aids for people with language impairments and improved brain–computer interfaces, though the work also sharpens concerns about mental privacy and consent. (Nature)

Real-Time Mandarin BCI Decodes Tones into Text

A Shanghai team reports a milestone for non-English brain–computer interfaces: real-time decoding of spoken Mandarin, a tonal language whose pitch contours change word meaning. Working with NeuroXess, researchers at Fudan University trained a system on a 43-year-old epilepsy patient who repeated ~400 Mandarin syllables over two weeks while electrodes recorded cortical activity. The decoder converted neural signals into on-screen Chinese characters at ~50 characters per minute with ~70% accuracy—slower than normal speech but a first for online Mandarin text generation. Unlike imagined-speech BCIs, the participant read prompts aloud, providing acoustic timing cues. Results build on the group’s 2023 intraoperative work and parallel Westlake University’s offline decoding advances. Next steps target stroke/ALS patients and wireless implants to boost speed, accuracy, and long-term usability. (Science)

Robotic Swimmer Uses Electromagnetic Fins to Match Fish Agility

Engineers have unveiled an underwater robot propelled by a flexible electromagnetic fin that brings fish-like speed and maneuverability much closer to reality. The prototype reportedly cruises at 405 millimeters per second—about 1.66 body lengths—and turns within a radius under one body length, thanks to a fin that oscillates under electromagnetic actuation rather than traditional motors or propellers. That design reduces mechanical complexity while offering precise, rapid control, with obvious appeal for ecological surveys and infrastructure inspections where quiet operation and agile motion matter. The work highlights how bioinspired actuation can outperform conventional thrusters in cluttered or sensitive habitats and could enable robotic swarms that glide through reefs or kelp forests without disturbing wildlife. (IEEE Spectrum)

MLPerf Shows AI Model Growth Outpacing Hardware Gains

Fresh MLPerf training results underscore an uncomfortable trend: AI model sizes and training demands are accelerating faster than hardware improvements can keep up. While GPUs and AI accelerators keep setting benchmarks, the newest submissions reveal diminishing returns per dollar and watt compared with the surging compute appetite of frontier models. That mismatch is already reshaping datacenter economics, energy planning, and research priorities, pushing teams toward algorithmic efficiency, model compression, and hybrid training strategies. Expect more investment in networking, memory bandwidth, and software tooling that clusters errors and prioritizes the fixes that move the needle fastest. In short, better chips help—but smarter optimization may matter more in the near term. (IEEE Spectrum)

China’s “Underwater Ghost” Jellyfish Robot Prioritizes Stealth

A biomimetic jellyfish robot developed at Northwestern Polytechnical University aims to ghost through the deep sea with jellyfish-like strokes, ultra-low power draw (~28.5 mW), and minimal acoustic signatures. Built around electro-hydraulic “muscle” actuators and hydrogel electrodes, the robot’s translucent body and propulsion mimicry are designed for covert observation, ecological monitoring in sensitive zones, and fine inspection of subsea infrastructure—use cases where traditional propeller-driven ROVs are too noisy or bulky. Researchers say the system’s low energy budget and high maneuverability could enable longer deployments, especially in swarms. It’s a reminder that soft, compliant designs are increasingly the go-to for ocean robotics where stealth, safety, and bio-friendly interaction are paramount. (New Atlas)

Royal Navy Fits Quantum Clock to Robotic Sub for Precise Nav

The UK’s Royal Navy has installed a quantum clock on its XV Excalibur robotic submarine, an upgrade meant to sharpen underwater navigation when GPS is unavailable. Quantum timekeeping allows the sub to dead-reckon with far lower drift, extending mission duration and accuracy for tasks such as mine countermeasures, intelligence gathering, and infrastructure inspection. Pairing a robotic platform with precision timing reduces reliance on surfacing or external acoustic beacons, making operations stealthier and more resilient. It’s also a signal that dual-use quantum tech is moving out of labs and into fielded systems, with maritime autonomy among the earliest beneficiaries. (New Atlas)

Soft “Exosuit Trousers” Aim to Unstick Astronauts from Stiff Suits

University of Bristol researchers designed soft robotic “exosuit trousers” to be worn under spacesuits, boosting mobility for astronauts who struggle with stiffness and fatigue during extravehicular tasks. The garment uses fabric-based artificial muscles and Kevlar anchoring to assist leg motion without adding heavy, rigid components. Beyond spaceflight, the team envisions support for people with mobility impairments or workers facing repetitive, load-bearing tasks. The approach reflects a broader shift toward textile robotics—distributed, low-profile actuators that integrate with clothing—promising comfort and safety improvements over traditional exoskeletons. Early tests suggest meaningful assistance while preserving range of motion, with future work focused on control algorithms and long-duration ergonomics. (phys.org)

---

---

AgiBot’s Human-in-the-Loop Factories Train Smarter Shop-Floor Robots

A new WIRED feature profiles AgiBot, a Chinese startup combining AI with human operators to rapidly teach robots diverse manufacturing tasks. The model blends large-scale perception and manipulation learning with on-site “labeling” and guidance from people, shortening deployment cycles and tackling jobs with too much variability for classic automation. The company’s pitch: general-purpose robots that can be re-skilled in days, not months, expanding automation beyond high-volume, highly structured lines. The piece situates AgiBot within China’s broader push to upgrade factories amid labor shortages and rising costs, and explores how human-in-the-loop methods may bridge today’s capabilities and the vision of truly adaptable, generalist robot labor. (WIRED)

Oxford’s “Brain-Free” Air-Powered Robots Synchronize Themselves

Oxford-led researchers demonstrated soft robots that move and even self-synchronize using only air pressure—no electronics, motors, or onboard computing. By carefully designing pneumatic circuits and chambers, the devices produce rhythmic, coordinated motions that can couple across units, similar to biological oscillators. This “fluidic intelligence” reduces complexity and cost while improving robustness in harsh environments where electronics fail. Potential applications include medical devices, grippers for delicate handling, and distributed morphing systems. The work, published in Advanced Materials, adds momentum to embodied-control approaches where mechanics and materials do more of the “thinking,” a promising pathway for safe, compliant automation. (EurekAlert)

Heavy-Duty Robotics Target Fusion Reactor Maintenance

China’s Hefei Institutes of Physical Science report advances in heavy-duty robotic joints and intelligent control tailored to maintain fusion reactors—tasks too risky for humans amid radiation, heat, and tight geometries. Published in Engineering Applications of Artificial Intelligence, the research optimizes joint parameters and control strategies for dexterity under load, aiming to speed inspection and remote replacement of reactor components. If adopted, such systems could reduce downtime and improve safety in experimental and future commercial reactors, where maintenance economics are pivotal. It also exemplifies how domain-specific mechatronics can unlock automation in extreme environments central to the clean-energy transition. (EurekAlert)

Artificial Neurons Bring Brain-like Signaling to Neuromorphic Chips

USC-led researchers built artificial neurons using ion-based diffusive memristors that more faithfully emulate how biological neurons process and transmit signals. The devices capture the chemical-to-electrical dynamics crucial for efficient, event-driven computation, potentially slashing power consumption for edge AI and embodied robots. By better modeling synaptic plasticity and temporal integration, these components could support more compact, low-latency control loops for mobile systems, including legged bots and drones. While still lab-scale, the approach complements spiking neural network research and could pair with sensor co-design to create tightly integrated perception–action stacks. (Science Daily)

Cooling Breakthrough Targets AI’s Heat—and Data Center Bills

Engineers detailed a passive evaporative cooling membrane that promises major heat-removal gains for next-gen electronics. As AI accelerators pack more transistors into tighter spaces, thermal bottlenecks increasingly cap performance and inflate energy costs; the new material system passively wicks and vaporizes coolant to sustain higher heat flux without bulky pumps or exotic fluids. If manufacturable at scale, it could shrink cooling footprints in servers and edge devices alike, enabling more compute per rack and extending hardware lifetimes. For robotics, more efficient onboard cooling could free power budgets for perception and planning, especially in compact, sealed systems. (SciTech Daily)