Bolei Deng's particles are dumb. Each one is a sliver of acrylic with thin tentacles, no sensor, no chip, no battery. Put twelve of them on a vibrating plate and they do something surprising: they form a solid block, then liquefy, then scatter like a gas. The behavior isn't programmed. It's geometry.
Deng, an assistant professor in Georgia Tech's Daniel Guggenheim School of Aerospace Engineering, and Xinyi Yang, a PhD student in his lab, published the work in Advanced Intelligent Systems in July 2025, alongside MIT CSAIL co-authors William Freeman and Wojciech Matusik. The paper appears on the journal's cover. The setup cost roughly $9,000 in acrylic, a Modal Shaker, and a vibration plate, no semiconductor anywhere.
The mechanism is elegant in the way old engineering tends to be. Each particle has flexible tentacle arms cut from a sheet of acrylic by laser. When two particles bump into each other, the tentacles latch, a bistable lock like a toggle switch clicking into one of two positions. The latched state stores elastic tension in the acrylic. When an external vibration source shakes the plate at the right frequency (100 hertz, 0.1 millimeter amplitude, producing accelerations that exceed gravity at roughly 39.5 meters per second squared), the interaction energy landscape flips from bistable to monostable. The lock releases. The particles push apart. With no central controller, no code, and no communication between units, the swarm reorganizes itself.
"In our system, you just send the vibration," Deng told Georgia Tech's research feature. "In space, once you build something, you need an astronaut or a robot to change it."
The swarm exhibits three collective phases. In the liquid phase, particles drift separately. In the solid phase, they interlock and lock as a block, strong enough to resist forces that would scatter the liquid assembly. In the gas phase, all particles possess kinetic energy and repel each other at relative velocities up to 974 millimeters per second. The transition is controlled entirely by vibration frequency and amplitude, not by any particle-level decision.
Each unit can be very dumb and follow simple rules, Deng said. But when you combine enough of them, a sort of intelligence begins to emerge.
That word, intelligence, is doing real work in the team's framing. The paper's title calls them "electronic-free particle robots," and the researchers repeatedly describe the design as encoding behavior. Change the geometry of the tentacles, and the swarm does something different. Surrogate simulation tools, developed to guide inverse design of particle geometries, let the team predict what shapes will produce what collective behaviors. The design is the algorithm.
"The intelligence is not programmed in, it is built in," Yang said. "Change the geometry, and you change what the swarm does."
There are limits. The paper notes a roughly 3.5-times efficiency gap between simulation and experiment, a gap the team attributes to fabrication imperfections, friction modeling, and the difficulty of precisely controlling the tentacle geometry at small scales. Simulations scaled up to 150 particles show robustness and fault tolerance increasing with swarm size, but the experimental maximum is twelve. Microscale particles have been fabricated using a Nanoscribe 3D printer with IP-S resin, demonstrating scalability from the width of a human hair to 1.5 inches, but the mechanical testing at those smaller sizes remains limited.
The team has also designed 3D particle robots with up to 162 tentacles, though the experiments described in the paper focus on planar designs with 32 tentacles. Sequential deployment, two particle types with different critical friction coefficients each responding to a different vibration amplitude, suggests programmable multi-behavior systems are within reach.
The applications the team envisions are either compelling or speculative depending on your tolerance for demos. Yang told TechXplore that particles could explore vessels no camera or catheter can reach, spreading into parts of the body otherwise invisible. Deng's space argument is more concrete: send a vibration to reconfigure a structure, no astronaut or robot required. Neither is close to a device. The gap between a twelve-particle acrylic demo and a medical or aerospace system is substantial, and the paper does not pretend otherwise.
What the work actually demonstrates is something more fundamental: mechanical computation through geometry, a swarm that behaves like something far more complex than the sum of its parts because the shape of each part encodes the rules. Whether that qualifies as intelligence depends on what you think the word means. The particles don't know anything. They don't decide anything. But they do something that looks like deciding, and that gap, between what a thing does and what a thing thinks, is exactly where robotics gets interesting.
"We're still just scratching the surface of what's possible when you let the design do the work," Yang said. That's either a genuine insight or a press-release line, and it's too early to know which.
† Add attribution (e.g., 'according to the researchers' or 'Deng said') or add footnote: '† Source-reported; not independently verified.'