Robots that feel, listen underwater, and lie less.

A prosthetic hand learns to feel how hard things are

The last time you reached into a bag and felt for your keys without looking, your fingertips were doing something a prosthetic hand cannot — yet.

Think about gripping a paper cup of hot coffee. Your fingers automatically adjust — not too tight, not too loose — based on constant feedback about pressure and texture. People using prosthetic hands lose that channel entirely. A research team working with the Hannes prosthetic hand, developed at the Italian Institute of Technology in Genoa, has built a system that starts to address this. They lined the hand's fingers and palm with 64 flexible sensors made from a material called PVDF — polyvinylidene fluoride, a plastic that generates a tiny electrical signal when squeezed or bent. A compact neural network reads those signals in real time and classifies what the hand is touching: soft, medium, or firm. In lab tests, it got this right about 91% of the time. The response took 0.21 seconds — roughly a quarter of the time it takes you to blink. In a separate offline test with nearly 12,000 measurement windows, accuracy climbed to 94.6%. They also tested the feedback direction: a small electrical pulse delivered to the user's skin to signal what the hand is 'feeling.' The system correctly located contact — finger versus palm — about 95% of the time. Here is the catch: every one of these tests happened in a controlled lab, gripping defined objects in defined ways. A wet glass, a crumpled bag, or a handshake is messier. The 64-channel sensor array needs to shrink considerably before it fits into a device someone wears all day. And there is no long-term wear data yet. This is a solid engineering result — not a finished prosthetic.

An underwater robot maps coral reefs by listening for fish

Finding the busiest corner of a reef, at night, without a map, on a single battery charge — that is roughly what this robot had to do.

Coral reefs are not uniformly alive. Some patches teem with fish; others are quiet. Knowing where the hotspots are matters for conservation — but surveying them by hand, with divers, is slow, expensive, and limited in range. A team whose work appears in Science Robotics deployed an autonomous underwater vehicle — a torpedo-shaped robot — at two reef sites in the US Virgin Islands: Joel's Shoal and Lameshur Bay. The robot carried two sensors: a downward-facing camera and a hydrophone, which is simply an underwater microphone. It also built 3D maps of reef rugosity — how bumpy and complex the seafloor structure is — because rough, complex terrain tends to shelter more life. Here is how the three signals combined: the camera fed video into a YOLO object-detection model — the same class of software used to spot pedestrians in autonomous vehicle footage — to count fish. The hydrophone picked up the clicks, grunts, and chirps that reef fish make when they aggregate. And the rugosity maps provided structural context. Together, the three signals produced heatmaps showing where biological activity was highest. The robot also demonstrated autonomous homing, returning to a target location after a survey on its own, across nine successful trials. The honest limitation here: what I have is the dataset deposit accompanying the paper, not the full paper text. Specific accuracy numbers — detection rates, false-positive rates — are in the Science Robotics publication, not available to me here. Two reef sites is a promising start; it is not yet evidence the system generalises to reefs globally.

One researcher's attempt to map the physics of AI hallucinations

Why does an AI chatbot confidently tell you a made-up citation? A solo researcher thinks they have found the culprit — and named it the 'Grammar Police.'

Language models hallucinate — they produce false information stated with full confidence. We know this happens. We are much less clear on why, mechanically, inside the model. A researcher working independently — not affiliated with a lab, not peer-reviewed — has published what they call 'Project Aletheia': a framework of seven claimed laws describing how hallucination works, using the vocabulary of physics. The central idea is intuitive: inside a language model, some processing layers retrieve facts, and others enforce grammatical structure. The claim is that the grammar-focused layers actively suppress factual recall — a bit like a meticulous copy editor who keeps smoothing sentences even when doing so strips out the original meaning. Using specific processing units called attention heads, identifiable at particular layers of the network, this suppression is claimed to explain up to 70% of factual errors. The researcher also claims that prefixing prompts with code-style symbols — comment markers like // or # — partially bypasses the suppression and improves factual accuracy. Now the part that matters most: I cannot recommend taking these specific numbers seriously yet. This is a solo investigation with no institutional review, no statistical significance testing, and no error bars on most of the claimed 'laws.' The primary testbed is GPT-2, a model from 2019 that is tiny by today's standards. Whether any of this extends to the models you actually use is unconfirmed. What is genuinely useful here is the vocabulary and the framing — fact retrieval and grammatical suppression pulling against each other is a real and testable hypothesis. Someone should run the proper experiments.

Three stories, and they are not as different as they look. Two of them — the prosthetic hand and the reef robot — are about AI leaving the screen and operating in physical space: on a body, underwater, in an environment where a wrong answer has a real consequence. The third asks whether we can trust AI's outputs at all, even on a screen. That is the tension running through all three today. The prosthetic hand is impressive precisely because its error rate is low enough to mean something for a real person gripping a real cup. The reef robot only helps conservation if its hotspot maps are accurate. And the hallucination paper, for all its methodological weaknesses, is pointing at a question that becomes more urgent the moment AI leaves the controlled setting. The pattern: we are building AI for physical, high-stakes contexts faster than we are building the tools to verify its reliability in those contexts. That gap is the story.

For the prosthetic hand work, the next meaningful step is long-term wearability trials — watch for clinical studies coming out of Italian Institute of Technology or partner hospitals in the next year. On hallucination, the International Conference on Machine Learning (ICML) runs in July 2026 and typically brings interpretability papers that will either support or challenge claims like Aletheia's — that is the moment to check back. The open question I would most want answered: does grammatical suppression of facts scale up with model size, or shrink? Nobody has a clean answer.

• Wikipedia: Piezoelectricity — background on how PVDF sensors work
• Wikipedia: Hallucination in artificial intelligence
• Science Robotics — the journal publishing the coral reef AUV study