A dairy farmer wrote to a university asking for help improving milk production. A helpful theoretical physicist responded by visiting the farm to gather data and analyze the problem. They then returned to the university to perform calculations and find a way to improve the farmer’s situation. A couple of weeks later, the physicist calls the farmer and says, “I have a solution, but it only works for spherical cows in a vacuum.” This classic joke is meant to illustrate the inherent absurdity of working out solutions to complex real-world problems using simplified parameters. Fortunately we no longer have to work out equations with pen and paper, and that has allowed MIT CSAIL researchers to find a way to quickly and efficiently design and control soft robots for specific tasks.
Designing robots and controlling them has traditionally been akin to working out a physics problem on some scratch paper. It’s relatively easy to work out the problem when your robot is rolling on solid wheels across a flat surface. But each new factor you introduce makes the problem more difficult to solve. Moving a soft, flexible robot across uneven and loose terrain is quite difficult. How much will it flex with each actuation? How much grip does its surface friction provide? How do you accommodate for loose sand in one area and hard stone in another area?
Those problems are nearly impossible to work out without extremely intensive processing. As Andrew Spielberg, the first author of this research paper, explains, “Soft robots are infinite-dimensional creatures that bend in a billion different ways at any given moment.” The trick then is to simplify the problem so that it can be calculated without the use of a supercomputer. However, in order avoid the “spherical cow,” the simplified problem still needs to represent the real world with reasonable accuracy.
The solution is to reduce those “infinite dimensions” down to a low-dimensional model with a similar output. In the past, that was done with a kind of manual trial-and-error approach in which a potential low-dimensional model needed to be evaluated for accuracy. This new material point method (MPM) technique automates the process with machine learning by simulating 3D-pixels (called voxels) of the robot’s body in a feedback loop. In testing, this model was able to successfully train robots in 75 times fewer simulations than those trained using traditional models. The next step is to create a full design pipeline for soft robots in the real world, and hopefully those robots won’t end up looking like spherical cows.