A team of researchers at Columbia University, funded in part by the Defense Advanced Research Projects Agency, have developed “machines that can grow by consuming other machines.”
Video of the experiment shows tubular robots that move by extending their shafts to inch along the ground. As the tubes gather, they connect and form into more complex shapes like triangles and tetrahedrons. With each piece consumed, the whole moves faster and with more elegance.
“AI systems need bodies to move beyond current limitations. Physical embodiment brings the AI into the messy, constraint-rich real world—and that’s where true generalization has to happen,” Phillipe Martin Wyder, lead researcher on the project, told 404 Media.
The researchers said the experiment was done with a view towards developing a “body” for AI. The idea is to give artificial intelligence a form that can grow, heal, and change similar to a biological body. They published their research in Science Advances under the title “Robot metabolism: Toward machines that can grow by consuming other machines.”
For the experiment, the researchers designed what they called truss links: “a simple, expandable, and contractible, bar-shaped robot module with two free-form magnetic connectors on each end.” Each truss link is almost a foot long when fully contracted and weighs more than half a pound. When the Links move individually they look like plastic worms inching across the ground, but their motion becomes more fluid and interesting as they gather to each other, forming complex shapes that allow them to move faster.
Right now, the truss links are controlled by a human on a keyboard and not artificial intelligence. “It’s not AI-controlled yet, but that’s partially the point: this architecture is a step towards future AI-controlled self-assembling physical systems,” Wyder said.
Wyder and his team controlled the truss links remotely and ran the robots through several obstacle courses. Some of the motions of the machines were preprogrammed with specially designed loops with names like “ratchet crawl” and “tetrahedron topple” that the researchers could activate with the push of a button. “There’s no autonomous AI running in the loop yet, but that’s the direction we’re heading,” he said.
Wyder said that giving AI a body was in its very early stages. “Miniaturization is also on the table—more links, smaller size, finer resolution,” he said. “But I don’t believe a single platform will suit every task. Deep-sea robots, Mars colony builders, assistive home systems—they’ll need different form factors. The deeper idea here is the metabolic principle, not just the physical design.”
Human consciousness happens at the point where the mind and body interact. A person is not just the thoughts in their head, but also how they react to their environment with their body. All that stimuli shapes our thoughts. Wyder and his team are seeking to, eventually, recreate this phenomenon for AI. The research is exciting, but it’s also very new and there’s no way to know how it’ll play out in the long term.
This need not be a world where AIs are stuck in human-like bodies. He pointed to previous research out of Sweden that used a swarm of robots to form furniture on demand. If such a system were to break, we should not expect the average person to be able to replace the part. But what if the system could order a replacement part and repair itself?
“For this vision to become a reality, we must build robot systems that are intelligent in a way that allows them to keep track of their changing morphology,” Wyder said. “When the idea of modular robots first surfaced in the late 80s this was unthinkable, but I believe that our recent progress in machine learning could allow for intelligent, modular self-assembling machines.”
He also acknowledged there are dangers here. “With our current robots, the worst-case risk is probably a pinched finger. But yes, autonomy plus embodiment demands careful consideration of all the risks. Most robots today still struggle with navigation and manipulation. They’re far from being autonomous agents in the wild, but rather need our care,” he said.
Wyder also said that he doesn’t consider the ethics of this work as an optional part of the research. “Malicious use of robotics is a broader concern and not unique to this platform. Like any powerful technology—nuclear, biotech, AI—governance matters,” he said. “I don’t think this class of robot poses near-term risks, but that doesn’t mean ethical foresight is optional. We have to think about it so we can get it right.”
The researchers will build on this work and that one direction is teaching robots how to exploit environmental factors. “Imagine a climber choosing which rocks to grab—robots need that same affordance awareness,” he said. “We’re working on how robots can reason about their environment and use it to drive reconfiguration or mobility.”
Along with the paper, the researchers have a GitHub and Zenodo that contain the CAD and mesh files, firmware, software, and simulation code for the truss links. Anyone, if they so desired, could build their own bundle of robot-devouring-robots.
