Researchers at ETH Zurich's Soft Robotics Lab taught a robotic hand to use its five fingers as legs and locomote on its own. It moved untethered across 14 surfaces, recovered from falls, and supported its own weight while pressing keyboard arrow keys, scoring 29 correct presses out of 32.
The experiments used an off-the-shelf WUJI hand fitted with a Raspberry Pi Zero 2 W, IMU and battery, with control policies trained in Isaac Lab and transferred to the physical system.
Give a robot hand mobility and suddenly the hand itself walks over and starts using the keyboard. It has become its own little creature.
XA10C4's Ruler Drive turns the familiar trick of twanging a ruler over the edge of a desk into a four-voice automated instrument. Stepper motors change each ruler's effective length to set its pitch, while solenoids strike the rulers; the machine is controlled over MIDI and the published demo performs the notoriously fast piece Rush E.
Built from an aluminum frame and 3D-printed parts, it mechanizes a very simple school-desk noisemaker into a programmable musical robot.
They've automated the classroom trick of twanging rulers—and multiplied it by four. Bring this to school and you might get scolded by four teachers at the same time.
Shivani Deshpande's Neon-Net is a mixed-reality game played by swinging a real aquarium fish net at AR jellyfish seen through Snap Spectacles. An Arduino UNO Q, IMU and vibration motor are mounted on the handle, where the system samples the net's motion and predicts the swing to decide whether a catch lands.
The jellyfish react to the player: approaching can make them flee, successful streaks can spook the shoal, and harder swings reach farther while disturbing more nearby creatures. Instead of leaving the interaction entirely in augmented reality, the project insists that virtual jelly-fishing should still involve holding an actual net.
AR could have made the whole game hand-tracked and prop-free, but jelly-fishing apparently still demands a real fish net. The digital world is doing the work to preserve the physical joke.
Casio's 1995 Loopy game console includes a built-in thermal sticker printer for printing images from games. Throaty Mumbo's LoopyDOOM port makes that unusual hardware feature available to DOOM too: an option in the menu prints the current game frame.
DOOM itself runs on the real 16 MHz SH-1 hardware with roughly 512 KB of RAM at around 8–15 fps. Running DOOM on odd hardware is already a familiar challenge; this port goes a step further and treats the console's printer as a feature DOOM ought to support too.
Getting DOOM running was already enough, but then the port noticed the console had a printer and decided DOOM needed screenshot printing too.
Petoi's mini robot dog Quaddle can keep walking even when flipped upside down. In the video, three Quaddles are also fixed together with blocks so they behave like one larger robot, then perform a backflip in unison.
A tiny quadruped walking around without seeming to care which side is up already gives it a peculiar creature-like quality. Combining three units turns them into something resembling one larger quadruped that rotates as a single body, making the demo increasingly unconcerned with both orientation and robot count.
First it keeps walking upside down as if nothing happened; then three of them combine for a backflip. This robot dog seems remarkably unconcerned with either up and down or how many bodies it has.
Kazuya Shibata's Fully Automatic Disposable Chopstick Splitter 2 is designed to split disposable wooden chopsticks cleanly and evenly. It feeds the chopsticks in, centers and pulls the two halves apart, then uses a circular saw to cut the remaining joint and a belt sander to smooth the ends.
The machine includes suspension to accommodate thickness variations and a self-centering gripping mechanism. Shibata spent about a year developing it, while processing one pair of chopsticks takes roughly four minutes. A task normally finished by hand in a moment has grown into a complete miniature production line.
All this machinery exists to avoid one bad chopstick split, and by the end a circular saw and belt sander have joined the process. The disposable chopsticks are receiving remarkably premium treatment.
Strange Inventions Lab turned an iPhone 13 Pro case into a miniature marble machine. The case contains a track and an Archimedes-screw lift, so tilting the phone sends a ball through the course and raises it for another lap.
The creator describes it as an experimental design that needs a fairly specific angle and can occasionally jam. It uses 19 screws and 12 ball bearings to add a complete mechanical toy to an object normally designed to stay thin and unobtrusive.
Phone cases are usually designed to disappear around the phone. This one adds a whole marble amusement park instead.
kimu-ryo's “True Random Number Generator Using a Disc Tray v1” uses the opening and closing motion of a DVD drive tray to physically roll a die. Because the result comes from an actual physical event, it serves as a source of randomness rather than a software pseudo-random sequence.
The ProtoPedia page highlights the simplicity of controlling the tray, and a v2 is being developed to return the die automatically after each roll. A computer component has effectively been reassigned to the ancient computational technique of throwing dice.
Instead of making the random-number algorithm smarter, the solution is to make a DVD drive physically roll dice. That's a dramatic return to hardware.
Adafruit's John Park built One Key, a wireless mechanical keyboard containing exactly one key. It connects as a Bluetooth HID device and can act as a mute button, presentation clicker, macro launcher or other configurable control.
Despite having only one physical key, it can send characters, combinations such as Ctrl-Z, strings and media commands. A WebSerial interface also configures LED colors and sleep behavior. The input hardware has been reduced to almost nothing, while the configuration for deciding what that one button means has become impressively rich.
The keyboard has been reduced to one key, but the settings for deciding what that one key does keep multiplying.
Throaty Mumbo published a patch for the original Super Mario Bros. that lets two Famicoms communicate through Nintendo's old Famicom Network System.
The setup calls for two consoles, two network units, compatible flash cartridges and phone cabling. The live demo uses a telephone-line simulator, while the README notes that a real landline might work as well. It is a 2026 multiplayer experiment whose networking stack has traveled conspicuously backward in time.
To connect with a friend in 2026, the required hardware is basically “two Famicoms and a phone line.” The route to the future is going backward.
Raul Muñoz wanted smart-treadmill data from a rented machine that he could not open, modify, or attach sensors to. His solution leaves the treadmill untouched: a USB camera watches it from the side while an AI model on an Arduino VENTUNO Q estimates walking speed from video and broadcasts the result over Bluetooth FTMS to compatible fitness apps.
To train the system, Muñoz recorded treadmill sessions while labeling the displayed speeds, building a dataset across ten sessions. A restriction that ruled out a simple physical sensor therefore led to an entire local computer-vision pipeline sitting beside the machine.
When you are not allowed to attach one small sensor, apparently the compliant alternative is to have an AI stare at the treadmill continuously. The rule is obeyed with impressive force.
Steven Gong and fellow University of Waterloo engineering students built a robot that can retrieve a shirt, straighten it, fold it, and place it on a stack. Two human-operated leader arms control two follower arms while cameras record the process; the team collected about 200 demonstrations for training.
Early models either failed to begin the task or moved violently enough to break a 3D-printed gripper. The team eventually fine-tuned a robotics foundation model, then had to collect more data after moving the setup to a new building changed the lighting. By the next morning, the system could complete the folding sequence autonomously.
A chore that takes a person less than a minute required four robot arms, hundreds of demonstrations, a broken gripper, and an overnight model search. Laundry has excellent scaling properties.
Give a robot hand mobility and suddenly the hand itself walks over and starts using the keyboard. It has become its own little creature.