Most materials are predictable. If you push a piece of steel, bend a plastic rod, or heat a sheet of glass, they respond according to fixed laws of physics. They don’t remember the last time they were bent, nor do they “learn” to be more flexible in response to repeated stress. In the world of robotics, we solve this by adding a brain—a central processor running thousands of lines of code that tells every motor and joint exactly how to move.
But a research team at the University of Amsterdam is challenging that fundamental divide between “dumb” matter and “smart” software. They have developed a metamaterial that learns through physical experience, effectively embedding intelligence into the object’s structure. Instead of relying on a central controller to dictate movement, the material updates its own internal settings based on how It’s handled, allowing it to remember shapes and adapt its behavior without a traditional software script.
As a former software engineer, I spent years thinking of intelligence as something that happens in a CPU or a cloud server. The idea of “physical learning”—where the memory resides in the stiffness of a hinge rather than a line of code—represents a paradigm shift in how we build machines. It moves us away from the “command and control” model of robotics toward a more biological approach, where the body and the mind are one and the same.
The findings, published in Nature Physics, describe a system that doesn’t just follow instructions but evolves its physical response to the environment. By distributing the “smarts” across the entire structure, the researchers have created a prototype that can be trained, can forget, and can switch between different tasks—all while remaining decentralized.
The architecture of a “brainless” machine
The prototype looks less like a high-tech robot and more like a mechanical toy: a worm-like chain composed of identical motorized hinges connected by an elastic skeleton. While each hinge is equipped with a microcontroller, there is no “boss” computer running the show. There is no master program directing the chain to curve left or right.
Instead, each hinge operates on local data. It measures its own rotation, remembers its recent motion, and exchanges brief updates with its immediate neighbors. Based on this local conversation, each hinge adjusts how hard it “pushes back” against external forces. When the entire chain interacts, these local adjustments coalesce into a coordinated global shape.
In one demonstration, the researchers trained the chain to form the letters that spell “learn” (or “leren” in Dutch). The chain didn’t do this because it was told the coordinates of the letters; it did it because it had been physically “taught” that those specific configurations were the preferred states.
| Feature | Traditional Robotics | Learning Metamaterials |
|---|---|---|
| Control Logic | Centralized (CPU/GPU) | Distributed (Local Hinges) |
| Memory Storage | Digital RAM/Flash | Physical Stiffness/Torque |
| Adaptation | Software Update/Recoding | Physical “Training” Epochs |
| Failure Point | Central Controller Failure | Graceful Degradation |
Training matter through “muscle memory”
Teaching a material is fundamentally different from programming a computer. The Amsterdam team uses a process they call “epochs,” which functions similarly to how a human might practice a physical skill. To teach the chain a specific shape, the researchers repeatedly nudge it into the target configuration.
During these epochs, the microcontrollers update the torques applied at the hinges. This changes the local stiffness of the material, altering how forces travel through the chain. Over time, the material develops a form of mechanical muscle memory. When the chain recognizes a familiar input setup, it automatically settles into the trained shape because that motion has become the path of least resistance.
This process relies on a principle borrowed from machine learning called “contrastive learning.” By comparing the current state of the material with the desired target state, the system updates its internal “learning degrees of freedom.” This allows the metamaterial to be flexible: it can learn multiple shapes, store them as “multistable responses,” or even “forget” an old shape to make room for a new one.
Beyond the lab: A solution for the e-waste crisis
While the ability to make a mechanical worm spell words is an impressive feat of engineering, the broader implications are environmental. We currently live in a cycle of rigid hardware; when a device’s function becomes obsolete or its physical form is no longer suited for a task, we discard it.
The scale of this problem is staggering. According to the Global E-waste Monitor 2024, the world generated approximately 62 million metric tons (68 million U.S. Tons) of electronic waste in 2022, with only 22.3% documented as properly collected and recycled. This waste stream is growing by roughly 2.6 million metric tons per year and is projected to hit 82 million metric tons by 2030.
A material that can be retrained in the field offers a glimpse of a more sustainable design philosophy. If a robot can be physically taught to perform a new task—switching from gripping a specific tool to crawling through a storm drain—without needing a hardware overhaul or a complete software rewrite, the lifespan of the device extends significantly. This adaptability is critical for deployment in “messy” environments, such as sensor networks in wetlands or inspection bots in infrastructure, where sending a technician to replace a rigid device is impractical.
The road to adaptable matter
It is essential to maintain a realistic perspective on the current state of this technology. These are research prototypes. They still require power sources, motors, and electronics, all of which carry their own environmental footprint. Scaling these “brainless” systems from a lab chain to industrial-grade materials will require solving significant challenges in power efficiency and material durability.
However, the conceptual breakthrough is permanent. By proving that intelligence can be distributed through the physical properties of a material, the University of Amsterdam team has opened a door to “adaptable matter.”
Lead researcher Yao Du and his colleagues are now focusing on more complex challenges, including how these materials handle “noise and uncertainty” and how they can develop time-dependent behaviors. The next phase of research will likely move toward integrating these learning rules into softer, more organic materials to further blur the line between synthetic machines and biological organisms.
We invite you to share your thoughts on the future of “physical AI” in the comments below. Do you think adaptable hardware could actually stem the tide of e-waste?
