Josh Bongard, a researcher at the University of Vermont, discusses computer-designed organisms with collaborators Sam Kriegman, Doug Blackiston, and Michael Levin from the University of Vermont and Tufts University.
Bongard frames AI progress across three dimensions: task performance, the range of tasks an agent can handle, and embodiment. He argues that allowing an agent’s body to change during training could offer a path toward more general capabilities. Google’s seven-arm robot system, trained for four months with a 1.2-million-parameter neural network, improved at grasping unfamiliar objects, but kept the same morphology and task.
Bongard traces morphology-changing robots to Karl Sims’s virtual creatures, whose forms evolved through selection for movement. Hod Lipson and Jordan Pollack later used 3D printers to build physical versions, while work by Nick Cheney and colleagues evolved soft robots from voxels with different material properties. Compositional Pattern Producing Networks, or CPPNs, encode patterns that determine a robot’s shape and materials.
Bongard and his collaborators adapted voxel-based evolution to design organisms from African clawed frog cells. Their simulations combine passive skin cells with heart muscle cells whose contractions are assigned random phase offsets, so evolution must find shapes and tissue layouts that produce directed movement despite unpredictable activity. One evolved design used two large rear legs and muscle tissue patterned along its underside.
At Tufts, microsurgeon Doug Blackiston separates and assembles frog cells, layering skin and heart muscle tissue before shaping the structures by hand. The resulting xenobots are under a millimeter long, and a built version moved in a similar direction to its simulated design; flipping it over stopped the movement. Bongard also describes groups pushing small pellets into piles, a hollow xenobot carrying an object, and xenobots healing after damage.
Bongard says future work includes automating construction, gaining more control over xenobot behavior, and adding cell types, sensory organs, and actuators. The xenobots described have no brains, though he says nervous tissue could be added. Michael Levin is interested in using them to investigate how cells communicate through bioelectric signals.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Dr. Josh Bongard, a roboticist at the University of Vermont, discusses xenobots, AI designed biological machines, and ways to bring robotics and synthetic biology into K–12 education with educators at Dynamic Landscapes 2022.