WorMa Worm Robot Shifts Water Ballast to Crawl, Climb and Swim

A robot that looks like a joke but works

The robots getting the most attention this month are expensive humanoids and multibillion-dollar AI models trained to run them. One of the more widely shared machines, though, is a 50-centimeter segmented worm with pool-noodle foam on its sides and two latex balloons full of water at either end. It's called WorMa, and it comes from NYU's Tandon School of Engineering. Over the past few weeks it has moved from a peer-reviewed paper into the general technology press, and part of the appeal is how odd it looks. Its creators, assistant professor Nana Obayashi and Ph.D. student Daniil Filimonov, picked the name by combining "worm" and "mass"2. New Atlas summed up its appearance plainly: it looks like a giant worm, built from five connected segments and four joints5.

The look is funny, but the engineering is serious. WorMa makes the case for the opposite of what most of the robotics industry is betting on right now: give the body more of the work, and leave less of it to the model.

What the team actually built

The reporting agrees on the basics. WorMa is a planar robot made of five rigid 3D-printed links joined by four rotational joints. It measures 50 cm long, 13 cm wide and 10 cm tall, and weighs 1.063 kg7. Four servos pass a traveling wave down the body, silicone pads underneath give it grip in one direction, and foam disks keep it afloat once it's in the water. An ESP32 microcontroller runs the servos and the pump2.

The main idea is the ballast. A 12-volt peristaltic pump in the middle segment moves water back and forth through silicone tubing between latex balloon tanks in the head and tail7. According to the paper, about 300 grams of water can be moved, roughly 28% of the robot's mass7. One outlet put the figure at nearly 350 grams2, but the published research and most other coverage say 30035. Draining one tank and filling the other takes a little under 110 seconds2.

The authors describe moving this mass as the robot's sole way of adapting to terrain. There are no extra actuators for reshaping the body and no switching between gaits7. That's what makes the work interesting. The robot uses the same wave-like motion everywhere. Only its center of mass changes.

The numbers behind the stunt

Across outlets, the performance figures match the paper. On a 19.5-degree slope, the head-heavy setup was the only one that climbed. It also lowered cost of transport by at least 33% compared with the other setups tested73. In water the logic reverses. With the ballast in the tail, WorMa swam up to 26% faster and 52% more efficiently than when the weight sat in the head3. One outlet recorded a peak speed of 9.6 centimeters per second2.

Steps are where the idea is clearest. WorMa comes up to a step with water in its head for traction, then pumps the water to the tail so the lighter front end can lift. Once the head catches the edge, the water goes back forward so the robot can pull the rest of its body up. This let it climb steps up to 15 cm high35. One account said a robot with a fixed mass distribution couldn't manage even a four-centimeter step2. Leaving the water works the same way: weight in the head gives the front pads enough grip to drag the body up a slope, while a heavy tail or an even split leaves it stuck at the waterline2.

The paper calls this the first demonstration of full land-to-water transitions by an undulatory robot7. That claim deserves attention, because snake- and worm-style robots have been good crawlers for years.

Embodied intelligence versus foundation models

WorMa's timing is what makes it worth more than a curiosity piece. It went public on September 144, during a stretch when the humanoid industry's attention is fixed on ever-larger robot foundation models.

Look at the recent releases. NVIDIA put out Isaac GR00T N1.7 as a generally available, commercially licensable vision-language-action model for humanoid skills. It pairs a vision-language backbone with a diffusion transformer that generates continuous actions, and its pretraining added 20,000 hours of human video11. A startup called Nature Will announced Natura-1, a 50-billion-parameter mixture-of-experts model for controlling dexterous hands, pretrained on more than a million hours of egocentric human video17. Figure has signed a deal for up to 100,000 Nvidia Vera Rubin GPUs to train its Helix model, starting with a $3.5 billion commitment17. Skild AI raised $1.4 billion in January for a robot-agnostic "brain"17. One industry roadmap describes humanoid development in late 2026 as a systems-engineering discipline that has to design body, control, data, foundation model and safety together15.

Obayashi explains WorMa in very different terms. She told the student newspaper that the robot relies on "embodied intelligence." In her words, the body itself handles most of the computation needed to adapt to its environment and stay robust in it4. The team also cites fish swim bladders as the inspiration for the mass-transfer system4. Building the robot took four months and about $3004.

I don't read this as a rejection of foundation models. It shows that a lot of the adaptability the industry wants to train into neural networks can be built into the hardware. WorMa doesn't have to recognize a slope and plan a whole new way of moving. It just moves its weight, and its contact forces with the ground change as a result. The paper says this works alongside undulation, because the gait already shifts mass along the body. Pumping water adds a slower layer of adjustment on top without changing the gait7.

Where the coverage diverges

Outlets mostly agree on the specs and differ in how they frame them. The trade and enthusiast press lean on the spectacle: a giant worm that swims and climbs stairs56. Interesting Engineering presents it as a way to simplify amphibious robots by swapping several specialized mechanisms for one adjustable system3. The student paper is the most cautious. It reports that the current version drifts in water and is too slow to be practical, and that more work and funding are needed before it can be deployed4.

That caution matters, and other details back it up. The paper notes lateral drift during the swimming trials, which the researchers had to correct for with regression when calculating speed7. A human still decides when the water moves2. Nearly two minutes to move the ballast is slow for any robot that has to react to changing ground. Real autonomy is still ahead. The team is working on faster pumps and sensor-based control so WorMa can detect terrain changes and shift its weight automatically3. The lab is also building a version with an onboard "brain" that processes its surroundings in real time, plus a diving mode for inspecting waterbeds without disturbing wildlife4.

Why it matters for the humanoid race

That planned upgrade is where WorMa's story meets the foundation-model debate. Once WorMa gets sensors and onboard intelligence, its controller will have a simpler job than a humanoid policy does. It won't be coordinating dozens of joints. It will be choosing among a few mass configurations: head-heavy, balanced or tail-heavy, which the paper treats as distinct settings7. When the body is smart, the brain can be small.

This matters more as the economics of robot foundation models shift. General Robotics CEO Ashish Kapoor has argued that general-purpose frontier models and robot-specific foundation models will merge. In that view, value moves to post-training for specific jobs, and the middle layer becomes a commodity17. If that's right, competitive advantage may come back to hardware: bodies that make tasks easier to learn to begin with. WorMa was built for a narrow job, monitoring coastlines and wetlands and inspecting flooded infrastructure, and Obayashi says its value is a robot that "doesn't need to be redesigned every time the terrain changes"3.

There's a wider lesson about where coverage goes. This fall's biggest viral robot moments have often featured humanoids, such as the Polish robot Edward Warchocki, which was filmed shouting at wild boars in Warsaw24. A small team spending $300 and four months is far from Figure's GPU budget. Still, WorMa shows that a well-designed body can handle problems that are otherwise left to massive models trained on huge datasets. The worm looks absurd. The design idea behind it holds up.