Roboticists often imitate nature to create humanoid robots for housework, worm machines for crawling in tunnels, and four legged devices for running and jumping that look like cheetahs. But they usually design an animal like robot body first, and then train artificial intelligence to control it. However, in biology, the body and brain co evolved to deal with complex tasks. Therefore, some researchers are borrowing a page from nature to design intelligent adaptive robots.
In the latest example, evolutionary roboticists at MIT have created a virtual environment in which algorithms can design and improve the physical form and controller of soft robots so that they can evolve at the same time. In this digital space called “evolutionary Gym”, the algorithm can develop robots for more than 30 different tasks, including carrying and pushing blocks, back crossing, climbing obstacles and climbing shafts. When MIT researchers used their own algorithms in their programs, software developed robots that were more effective than humans for each task.
Jagdeep Bhatia, an undergraduate in MIT’s computer science and Artificial Intelligence Laboratory, led the work. He said: “the goal in the future is to undertake any task and say ‘design me the best robot to complete this task’.”. He presented the study at the conference on neural information processing systems on December 9.
The evolutionary gym relies on two algorithms that bounce back and forth. First, the design optimization algorithm “generates a series of random robot designs,” Batia said. The algorithm creates each soft robot by combining up to 100 independent building blocks, which can be rigid or flexible, and can move vertically or horizontally. Then these piecemeal designs enter the control optimization algorithm, which generates a “brain” for each robot to perform a given task. The controller calculates when and how much each block should be activated, for example, the distance and frequency of horizontal moving block activation, so that they work together to move the robot as needed. Next, various robot designs try to assign tasks in the evolutionary gym, while the control optimization algorithm measures their performance and returns the scores to the design algorithm.
Enter the principle of evolution. Bhatia said that the design algorithm abandoned the inappropriate configuration and “retained the most appropriate designs and changed them slightly to see if they performed better.”. From the design algorithm to the controller algorithm, the robot is tested in the evolutionary gym environment, and then returns to the design algorithm until the system converges to the highest score. The result of this process is design and control or the best combination of body and brain to complete the task.
Usually, this process produces familiar shapes. In order to complete the climbing task, the award-winning design evolved two arms and two legs to help it climb a shaft like an ape. The best handling robots look like a mixture of dogs and sticky shopping carts. But in most cases, the results were unexpected. They are not like devices designed by real animals or humans, but like things made of building blocks by a toddler.
Bhatia’s favorite task is that the robot must slide under randomly spaced tiles, and then drag an object on the top of the tiles while still under the tiles. In order to complete this work, the simulator designs a robot, which unfolds under the ceramic tile once, and then swings slowly to push the object onto it. This is a perfect example of the body and brain working together to take wise action.
This is the beauty of evolutionary robots, said Josh Bongard of the University of Vermont, who was not involved in the work. He said that the plan to replicate the body of robots from nature usually does not work, “because the evolution of dogs and humans fills a niche very different from the environment we are trying to introduce humanoid robots or canine robots.” Bangard explained that the aviation industry is a good example. “Early pioneers tried to make wing beating machines, but these prototypes failed,” he added. “Only when we make non bird machines can we make them fly.” Similarly, robot bodies generated by evolutionary algorithms usually look strange, but seem to work well in a given task.
Batia said others tried to design the body and brain of the virtual robot together, but they focused on simple tasks such as walking and jumping. “The most important thing in our work is the number of tasks we develop and the number of unique tasks,” he said.
Evolution gym is open source: Bhatia’s team created it to provide a benchmark platform on which any researcher can design and test their own algorithms and compare various methods. In previous work, teams often developed their own virtual environments for such assessments. The new digital space provides researchers with a common benchmark to measure the superposition effect of various algorithms. “It allows people to measure progress, which is very important,” said agrim Gupta, PhD in computer science. Students doing similar research at Stanford University. He recently published a paper on the development of intelligence not only through evolution, but also through experience. Bongard agrees, saying that the new M.I.T. simulator will allow the field of evolutionary robots to “move forward faster by clarifying which evolutionary robots work better than others.”
This evaluation is necessary because robots designed by evolutionary algorithms do not always work. For example, the M.I.T. algorithm cannot successfully design a robot for grasping and lifting. Batia said that this shows that there is still a lot of work to be done in designing truly intelligent robots, which makes standard platforms such as evolution gym more important for jointly promoting the development of robot design. As he said, “we are developing more intelligent artificial intelligence algorithms to create intelligent robots in real life in the future.”