Evolution Gym: A Design System that Can Evolve Robots

Imagine that you are running a race. To complete it, your body needs to be strong, and your brain needs to track the route to control your steps and prevent you from tripping. The same applies to robots. To complete tasks, they require both a carefully designed body and a brain or controller.

Engineers can use various simulations to improve the control of robots, making them more intelligent. But there are few ways to simultaneously optimize the design of robots.

Unless the designer is an algorithm.

Thanks to the advancement of computing technology, it is finally possible to write software that simultaneously optimizes design and control, a method known as collaborative design. Although there are established platforms to optimize control or design, most collaborative design researchers have to design their own testing platforms, which are usually very computationally intensive and time-consuming.

To help solve this problem, Jagdeep Bhatia, an undergraduate researcher at the Massachusetts Institute of Technology, and other researchers created a 2D collaborative design software robot simulation system called Evolution Gym. They presented the system at this year's conference on neural information processing systems. Now, they have also provided a detailed introduction to the system in a new paper.

"Basically, we attempted to create a very simple and fast simulator," said Bhatia, the first author of the paper. "Based on this, we established a series of tasks for these robots."

In Evolution Gym, 2D soft robots are composed of colored units or voxels. Different colors represent different types of simple components - either soft materials, rigid materials, or horizontal or vertical actuators. The result is that the robot is pieced together from colored blocks and moves in a video game like environment. Because it is 2D and the program design is simple, it does not require too much computing power.

As the name suggests, researchers structured the system to mimic the evolutionary process of organisms. It is not generating a single robot, but rather generating a population of robots with slightly different designs. The system has a two-level optimization system - an outer loop and an inner loop. The external loop is a design optimization. The system generates several different designs for a given task, such as walking, jumping, climbing, or grabbing something; The inner loop is used for control optimization.

Bhatia pointed out that the system will adopt each of these designs, and it will optimize the controller for a specific task in Evolution Gym. Then, it will return a score for each design to go back to the design optimization algorithm and say, "This is the performance of robots using better controllers.".

Through this approach, the system generates multiple generations of robots based on the "reward" scores for specific tasks and retains elements to maintain and increase these rewards. Researchers have developed over 30 tasks for robots to attempt, which can be classified as simple, moderate, or difficult.

"If your task is walking, in this case, you want the robot to move as quickly as possible within the specified time," said Wojciech Matusik, a professor of electrical engineering and computer science at MIT and the first author of the paper.

Researchers have found that the system is highly effective for many tasks and algorithm designed robots perform better than human designed robots. The system has come up with designs that humans can never achieve, capable of producing complex materials and highly effective actuators. Although the system had no prior knowledge of animals or biology, it independently came up with some animal like designs.

On the other hand, no robot design can effectively accomplish very difficult tasks, such as lifting and grabbing objects. Wolfgang Fink, an associate professor of engineering at the University of Arizona who did not participate in this study, pointed out that there may be many reasons for this, including the lack of diversity in the evolutionary population of program selection.

In addition, the simplification and 2D design of Evolution Gym are not yet suitable for adaptation into real-life robots. Nevertheless, Bhatia hopes that Evolution Gym can become a resource for researchers and enable them to develop new and exciting co design algorithms. At present, the program is open source and people can use it for free.

Prev: Scientists have developed intelligent window coatings that can cool or heat indoor environments according to the climate