At the World Humanoid Robot Games, a humanoid attempting a weightlifting event lifted the barbell, toppled forward and crashed into the judges' table. The failed lift was dramatic enough, but what followed made the scene look even more like a conventional sports broadcast.
Staff picked up the fallen machine, placed it on a stretcher and carried it away. From competition failure to post-incident removal, the sequence accidentally mirrored the familiar rhythm of an injured athlete leaving the field.
By the end, the robot is getting the full injured-athlete treatment. A stretcher adds an amazing amount of sports drama.
In Chicago's Lincoln Park, more than a dozen delivery robots clustered together on a sidewalk and appeared unable to move past one another. Coco said a deployment error caused multiple units to reroute at the same time, briefly creating the group.
Autonomous machines intended to make delivery more efficient ended up reproducing a familiar human systems problem: too many agents trying to use the same route at once.
Automate delivery far enough and robot society gets its own rush hour. We didn't need that part of human life reproduced.
Practicing badminton alone creates a simple problem: someone still needs to feed the shuttlecocks. This machine solves it by putting a compact Denso VS050 industrial robot arm in charge of shuttle delivery.
A pneumatic gripper picks one shuttlecock from a hopper and carefully feeds it between two high-speed spinning discs for launch. Even metal 3D-printed parts are involved, turning the everyday lack of a practice partner into something closer to a factory automation cell.
No practice partner? Bring in an industrial robot. That's an impressively large equipment budget for a very ordinary problem.
Palmimo DevKit is a tabletop AI robot that lets LLM tool calls map directly to physical actions such as walking, waving and turning its head. With six legs plus a neck, it has 21 degrees of freedom, along with a Raspberry Pi 5 and a round display face.
The warm wooden exterior is designed to feel approachable, and motions can be called from Python. Technically it is a physical-AI development kit; visually it can look like a small piece of furniture that decided to move around.
Inside it's a serious AI robot, but in a room it still looks surprisingly furniture-adjacent. That's a charming mismatch.
Making a robot feel alive is not only about faces or voices; whole-body motion can matter too. MOFU (MOrphing Fluffy Unit) explores that idea with a fluffy spherical body that expands and contracts. Developed by researchers at the University of Electro-Communications, it uses a Jitterbug geometric structure and a single linear actuator to change its height from roughly 210 mm to 280 mm.
In user studies, adding whole-body expansion, contraction and rotation made the robot seem more animate than locomotion alone. Rather than relying on a face, MOFU uses the rhythm of its entire body to suggest life.
Before giving the robot a face, they basically said “let's make the whole body breathe.” The fluffiness makes that strangely convincing.
Robot falls and crashes are now being treated as a subject worthy of asking why humans enjoy robot fails so much. The Wall Street Journal used clips from events such as the World Humanoid Robot Games as the basis for a broader analysis.
Rather than simply collecting funny accidents, the piece discusses slapstick, reassurance and anxiety about technology. Robot fails are shifting from isolated viral clips into something closer to a recognizable cultural genre.
We laughed at robot fails long enough that now there are essays explaining why we laugh at them. That's a healthy sign of genre maturity.
At the World Humanoid Robot Games 100-meter final, Tiangong Ultra recorded 8.64 seconds while several robots continued into protective barriers after crossing the finish line. Sprinting speed has advanced quickly; post-race stopping remains a separate challenge.
Reuters footage showed machines hitting walls, one requiring a fire extinguisher and others being carried away. Impressive running performance and unfinished stopping control appeared side by side at the same finish area.
An 8.64-second run is impressive, then the scene after the finish suddenly looks like an incident report. Speed and composure are clearly separate subjects.
A soft robot has been developed that, under infrared light, twists to store energy, kicks itself off the surface, lands and resets its shape on its own. The NC State University design uses a teardrop-shaped liquid crystal elastomer.
As long as the light continues, it repeats a cycle resembling charge → jump → reset. Adjusting the tail angle changes whether it crawls, jumps forward or hops nearly straight upward.
The energy cycle is beautifully engineered, while the result looks like a mysterious desk object that occasionally remembers to jump. Hard not to watch.
To push 3D printing down to roughly 150 µm scale, a microscopic printer was built that shines a 405 nm laser through the tip of an optical fiber and cures resin point by point. Created by Diffraction Limited, it moves the illuminated point much like an FDM printer moves its nozzle.
To improve resolution further, the resin is mixed with curcumin from turmeric as a UV absorber. The result includes a Benchy just 150 µm long and a Stanford Bunny that appears smaller than a human hair.
Push microfabrication far enough and turmeric somehow joins the materials team. That's an unexpectedly kitchen-adjacent breakthrough.
If an alarm sound is not enough, why not have the clock play a 720×720 video when it's time to wake up? The Video Alarm Clock uses an ESP32-P4-based Waveshare module and can store plenty of clips on an SD card.
The futuristic part comes with some very old-fashioned requirements: AVI with MJPEG and PCM, plus classic 8.3 filenames. The display feels modern; preparing the files suddenly feels like an earlier era of computing.
It's a video alarm clock from the future, quietly asking you to respect 8.3 filenames in the background. Nice little time warp.
Generative image models usually suggest large GPUs, but this project asks whether one can be squeezed into an RP2350 microcontroller with just 4 MB of flash. It produces 128×128-pixel faces at roughly one image every 20 seconds.
The model and inference code fit into the tiny memory budget, and output can even be displayed directly through a VGA adapter. Instead of scaling AI upward, the project explores how far it can be compressed downward.
While everyone else adds GPUs, someone is removing things until the model fits in 4 MB. It's satisfying to see the race run backward.
Robot mowers often struggle to trim right up against walls and lawn boundaries. The H510 Pro tackles that limitation with a mechanical arm that moves its main cutting disc outward by about 7.4 cm. The system was shown at IFA 2026.
EdgeSense identifies different boundary types, and when needed the mower physically reaches the blade toward the edge. Side guards and people-and-animal detection handle safety while the machine solves the last few centimeters of mowing in an unusually literal way.
Can't quite reach the edge? Extend the blade on an arm. It's such a wonderfully direct answer to a tiny geometric problem.
By the end, the robot is getting the full injured-athlete treatment. A stretcher adds an amazing amount of sports drama.