Machines Making Machines

Most of us have grown to be familiar with the idea of using prompt engineering techniques to coax a machine learning algorithm into producing poetry, artwork, or code. But what if we could write prompts that would produce something more tangible, like a complex, physical machine? A tool of this sort could radically change many fields, reducing development cycle time while increasing creativity and efficiency. And as it turns out, just such a tool was recently created.

Robots are very challenging to build because of the complex interactions that exist between the body structure, the actuators, sensors, and the environment. Because of these types of complexities, development cycles tend to be long and iterative. Engineers try something out, observe the results, then make tweaks until the system is working as expected. Not only is this process slow, but it is also limited by the creativity of the designers.

In an effort to eliminate these months- or years-long development cycles, engineers have turned to computer simulations for help with the design work. But traditionally, these simulations operate more or less by trial-and-error. As such, the problem space is massive when evaluating new designs, and extensive runtimes on supercomputing equipment is required. The process is simply too inefficient for much of any real innovation to be realized.

A team led by researchers at Northwestern University realized that trial-and-error is a dead end in the complex world of robot design. They developed a new method that instead leverages a gradient descent algorithm. Gradient descent is a critical algorithm involved in the training of virtually all of the important machine learning tools that we use today. Rather than blindly trying random values in the search for a solution to a problem, gradient descent instead progressively improves models as they approach an optimal state.

The model is lightweight and capable of designing a functional robot in about 30 seconds when running on a common, off-the-shelf laptop computer. Needless to say, that is dramatically faster than either traditional computer simulations or design by human engineers. It is also a highly creative tool. The model makes very few assumptions and starts with a minimalistic, randomly-generated body plan. A user then supplies a prompt, such as: “design a robot that can walk,” and the algorithm gets to work, morphing the design to match the user’s request. And with the help of the gradient descent algorithm, the design gets better and better with each iteration.

Using their tool, the team asked for a design for a walking robot. Within seconds, this produced the blueprint for a 3D-printable mold. The mold was used to cast a soft robot with pneumatic actuators and three legs. It may not have been pretty, and it is unlikely that any human designer would have ever arrived at a similar design, but it did prove to be effective. The little purple robot was demonstrated to be capable of walking at a rate of half of its body length per second (that is about half as fast as a typical human).

Interestingly, the researchers did not fully understand the design. The algorithm chose to, for example, punch a number of holes in the robot. They seem like they may just be unnecessary artifacts at first glance. But it was discovered that when they were removed, the robot could no longer walk.

In the future, the team hopes to see their system being used to design tiny robots that can travel through the human body to treat a variety of ailments. That may still be quite a long way off in the future, but there could be some more immediate benefits as this tool helps to enhance the creativity of robot designers, and improve the efficiency of their creations.