Robotics has spent decades trying to build a thumb-sized machine that can climb over rubble, carry its own power, and not tip over. Two labs on opposite sides of the planet gave up and hired cockroaches instead. One taught them to read terrain. The other handed them a syringe. Also this week: a pop star on a biotech board, and a ring that measures your blood sugar without breaking skin. Nobody here took the obvious route.
Table of Contents
🪳 We Gave Cockroaches AI, and They Started Climbing Better
RESEARCH
Every cyborg insect controller built so far has treated the insect as a chassis with legs. See obstacle, zap antenna, steer around it, repeat. Which works, in the sense that a person dragging a dog past a fire hydrant is also technically walking the dog. A team at the University of Osaka decided the roach might have opinions worth listening to.
Mochammad Ariyanto and Keisuke Morishima, working with Universitas Diponegoro, strapped a sensor backpack to cockroaches and ran a small neural network on it. The classifier sorts what is in front of the insect into four categories: flat ground, ascent, descent, and hole, and gets it right 92% of the time in offline testing. It then decides whether to steer at all.
That last part is the actual advance. When the system recognizes a climb, it stops issuing commands and lets the animal do the thing its lineage has been doing since the Carboniferous. The paper in Device reports shorter obstacle-crossing times, shorter total routes, and fewer electrical pulses delivered per run. Morishima frames the design goal as recognizing terrain "without compromising the insect's natural locomotor abilities", which is a polite way of saying the previous approach was fighting the roach.
Morishima's lab has been building these things for years, and the trajectory is consistent: every improvement involves the computer doing less. The best control algorithm for a cockroach turns out to be knowing when to shut up.
💉 The Cockroaches Have Syringes Now
RESEARCH & NEWS
Cyborg insects have spent twenty years being very good at finding people and completely useless afterwards. You locate a survivor under a collapsed floor slab, and then what? You radio it in and wait. Thang Vo-Doan's group at the University of Queensland asked the obvious follow-up: "Once they find someone, can they actually help?"
Their answer is the paraborg, and the platform is a Macropanesthia rhinoceros, the North Queensland giant burrowing cockroach, which at roughly 40 grams is the heaviest roach on Earth and can haul more than half its own body weight. Working with UNSW's medical robotics lab, the team fitted it with steering electrodes in the antennae and cerci plus a spring-loaded injector. A chemical reaction between citric acid and potassium bicarbonate generates a puff of carbon dioxide, and that pressure drives the drug through a 31-gauge needle. A tiny insect-mounted seltzer tablet, essentially.
The results reported in Advanced Science are honest about the gap between demo and deployment: 95% success on close-range injection within 15 centimetres of the target, falling to 72% once the roach has to navigate there first. Targets were silicone pads and pig skin, not people, and nobody has yet checked whether Bluetooth holds up through concrete and soil debris. Vo-Doan puts deployment at five to ten years.
To be fair, "your rescuer is a cockroach carrying a needle" is a hard sell for someone trapped in the dark. It still beats nobody arriving.
💄 Lady Gaga Is On A Biotech Board, And The Biotech Is About Skin
NEWS
Every celebrity eventually launches a skincare line. Very few of them come with a machine-learning platform, a tissue bank, and four years of stealthily figuring out why human skin dies in a dish.
Outer Bio came out of stealth in late August with $23 million from investors including Wing, Initialized, and Hawktail; Michael Polansky as CEO, 19 employees, and Stefani Germanotta, better known as Lady Gaga and Polansky's fiancée, on the board of directors. The pitch is not a cream. It is that donated human skin normally falls apart in about a week, while Outer Bio's YUNA platform keeps full-thickness human skin alive for roughly four weeks.
Four weeks is long enough to actually test something. The tissue keeps its structure and its immune cell function, so the company can run compounds against it, generating over 30,000 data points per compound-tissue pairing and more than 10 terabytes across 300-plus donors, then train models on its own data rather than the public literature. Polansky claims the loop now produces a new candidate every six weeks instead of every 18 months.
The strategic move here is quite interesting, and it is entirely about regulation. Skincare ingredients do not need FDA approval. Drugs do. So a platform with genuine pharma applications is aiming itself at the cosmetics aisle, where it can sell things this decade. That is either a shrewd read of the market or a slightly depressing verdict on how long the other path takes.
💍 A Ring That Measures Your Blood Sugar From Finger Sweat
RESEARCH
Your smart ring knows your heart rate, your sleep stages, and how badly you slept after the second coffee. What it does not know is a single thing about your actual chemistry. As UC San Diego's Tamoghna Saha puts it, commercial rings "lack molecular information about biochemical markers", which is the polite version of "your ring is a very expensive metronome."
Joseph Wang's lab built CHARM, described in Nature Communications, to fix that. The trick is getting sweat without sweating: an osmotic hydrogel pulls fluid out of the finger by pressure gradient alone, no exercise and no electrical stimulation, at a trickle of 100 to 150 nanolitres per minute. An electrochemical sensor array reads it. Everything, battery and wireless electronics included, fits in a normal ring.
It handles a panel of six: glucose, ketone, lactate, uric acid, vitamin C, and alcohol, four at a time because they share a reference electrode. Glucose tracked commercial CGMs with a mean absolute relative difference of 13.72%, and the other markers correlated at roughly 0.85 to 0.90 against blood meters.
Now the caveats, which are load-bearing. The validation cohort was three people, one healthy and two with type 1 diabetes. In the diabetic participants, the error rose to 17.05%, well off what you would want for insulin dosing. The battery lasts about 12 hours in the current prototype, and better water resistance is still on the to-do list.
Still: continuous ketones and glucose from the same finger, no needle. The concept works. The engineering is the boring bit, and the boring bit is where these things usually wither and die.
Four stories this week, and every one of them is a scientist quietly admitting the machine should do less. Osaka spent an entire paper teaching a computer when to stop giving orders. Queensland looked at two decades of rubble robot demos and went out and bought a bug. Outer Bio's model learns from skin instead of from papers about skin. Even the ring gets its sample by asking nicely, osmosis instead of an electric shock.
There is probably a lesson in here about engineering, or possibly about management. We have decided not to look directly at it.
Which one landed hardest? Hit reply; we read everything. And if you know someone who needs to hear that a cockroach may one day inject them with adrenaline, forward this along.
Keep questioning everything (especially anything with a backpack on),
P.S. Two independent labs on opposite sides of the world spent August making cockroaches more capable, and neither press release mentioned an off switch. And greetings from Finland, where the coming winter remains our most reliable pest control strategy.