Imagine that you are running a race. To complete it, your body needs to be strong, and your brain needs to follow the route, control your pace, and keep you from tripping.
The same is true for robots. To accomplish tasks, they need both a well-designed body and a “brain”, or controller. Engineers can use various simulations to improve the control of a robot and make it smarter. But there are few ways to optimize a robot’s design at the same time.
Unless the designer is an algorithm.
Thanks to advances in computer science, it is finally possible to write software that optimizes both design and control, an approach known as co-design. Although there are established platforms to optimize control Where design, most co-design researchers had to design their own test platforms, and these are typically very time consuming and computationally intensive.
To help resolve this issue, Jagdeep Bhatia, an undergraduate researcher at MIT, and other researchers created a co-designed 2D flexible robotics simulation system called Evolution Gym. They presented the system at this year’s Neural Information Processing Systems conference. They also detailed the system in a new document.
“Basically we tried to make a simulator that was really simple and fast,” said Bhatia, the lead author of the article. “And on top of that, we’ve built like a bunch of to-do’s for these robots.”
In Evolution Gym, 2D soft robots are made up of colored cells, or voxels. Different colors represent different types of simple components – either soft or rigid material, and horizontal or vertical actuators. The results are robots which are patchworks of colored squares, moving in environments resembling video games. Because it is in 2D and the program is designed simply, it does not need a lot of computing power.
As the name suggests, the researchers structured the system to mimic the biological process of evolution. Rather than generating individual robots, it generates populations of robots with slightly different designs. The system has a two-level optimization system – an outer loop and an inner loop. The outer loop is design optimization: the system generates a number of different designs for a given task, such as walking, jumping, climbing, or grabbing objects. The inner loop is the optimization of control.
The researchers found that the system was very efficient at many tasks and that the robots designed by algorithms performed better than those designed by humans.
“It will take each of these models, it will optimize the controller in Evolution Gym for a particular task,” Bhatia said. “And then it’ll return a score for each of those designs to the design optimization algorithm and say, that’s how well the robot performed with the optimal controller.”
In this way, the system generates multiple generations of robots based on a task-specific “reward” score, retaining the elements that maintain and increase that reward. Researchers have developed more than 30 tasks for robots to attempt to perform, categorized as easy, medium, or difficult.
“If your task is to walk, in that case you would want the robot to move as fast as possible within the allotted time frame,” said Wojciech Matusik, professor of electrical engineering and computer science at MIT and lead author of the article.
The researchers found that the system was very efficient at many tasks and that the robots designed by algorithms performed better than those designed by humans. The system came up with designs humans never could, generating intricate patchworks of materials and highly efficient actuators. The system also independently offered animal designs, although it had no prior knowledge of animals or biology.
On the other hand, no robot design could effectively accomplish the most difficult tasks, such as lifting and grabbing objects. There could be a number of reasons for this, including populations where the program selected to scale was not diverse enough, said Wolfgang Fink, an associate professor of engineering at the University of Arizona who was not. not involved in the project.
“Diversity is the key,” he said. “If you don’t have the diversity, you get great success quickly, but you are most likely stabilizing yourself in a suboptimal way. In the MIT researchers’ most effective algorithm, the percentage of robots that “survived” each generation was between 60% and 0%, gradually decreasing over time.
Evolution Gym’s simplistic 2D designs also don’t lend themselves to adaptation into real-life robots. Nonetheless, Bhatia hopes that Evolution Gym can be a resource for researchers and enable them to develop exciting new algorithms for co-design. The program is Open source and free to use.
“I think you can still gain a lot of valuable information by using Evolution Gym and coming up with new algorithms and creating new algorithms inside of it,” he said.