Robotics often copy nature, building humanoid robots for household chores, worm-like machines for crawling through tunnels, and four-legged gadgets that look like cheetahs for running and jumping. But usually they first design an animal-like robot body and then train an AI to control it. In living things, however, the body and brain develop together to cope with complex tasks. So some researchers are borrowing a page from nature’s playbook to develop intelligent, adaptable robots.
In the latest example, evolutionary robots at the Massachusetts Institute of Technology created a virtual environment where algorithms can design and improve both the physical shape of a soft robot and its controls so that they evolve at the same time. In this digital space called Evolution Gym, algorithms can develop robots for more than 30 different tasks, including carrying and pushing blocks, performing backflips, negotiating barriers, and climbing manholes. When the MIT researchers used their own algorithms in the program, the software developed a more effective robot than a human for each individual task.
“The future goal is to take on every task and say, ‘Design me an optimal robot to do the job,'” says Jagdeep Bhatia, a student at MIT’s Computer Science and Artificial Intelligence Laboratory who led the work. He presented the research results at the Conference on Neural Information Processing Systems on December 9th.
The Evolution Gym is based on two algorithms that jump results back and forth. First, a design optimization algorithm creates “a series of random robot designs,†says Bhatia. The algorithm creates each soft robot by combining up to 100 individual building blocks, which can be rigid or flexible and can move vertically or horizontally. These patchwork designs then go to the control optimization algorithm, which generates a “brain†for each robot that enables it to carry out a specific task. This controller calculates when and how strongly each block should be activated – for example, how far and how often a horizontally moving block is pressed – so that they all work together to move the robot as needed. Next, the different bot designs try out an assigned task in the Evolution Gym while the control optimization algorithm measures how well they are doing and feeds the results back to the design algorithm.
Enter evolutionary principles. The design algorithm throws out inappropriate configurations, “keeps the most suitable designs, tweaks them a little and sees if they work even better,” says Bhatia. This goes on as robots move from the design algorithm to the controller algorithm for testing in the Evolution Gym environment and back to the design algorithm again until the system converges to the highest score. The process gives the best combination of design and control – or body and brain – to get the job done.
Often times the process leads to familiar forms. For the climbing task, the winning design developed two arms and two legs that help him slide up a shaft like a monkey. The best baby carrier looks like a cross between a puppy and a squishy shopping cart. But in most cases the results are unexpected. Instead of resembling a real animal or device that a human would design, they look like something a toddler could have built out of blocks.
Bhatia’s favorite arose from a task that required the robot to slip under randomly arranged tiles and then drag an object over the top of those tiles while it was still underneath. For this job, the simulator designed a bot that unfolded once under the tiles and then slowly wobbled to push the object over them. It’s a perfect example of how the body and brain work together to act intelligently.
That’s the beauty of evolutionary robotics, says Josh Bongard of the University of Vermont, who wasn’t involved in the work. Copying robot blueprints from nature often doesn’t work, he says, “because dogs and humans have evolved to fill completely different environmental niches than we are trying to introduce our humanoid or dog-like bots.” Aviation is a good example, explains Bongard. “Early pioneers tried to make machines with flapping wings, but those prototypes failed,†he adds. “It wasn’t until we started building non-bird-like machines that we made them fly.” Similarly, robot bodies made by evolutionary algorithms often look strange, but seem to work well for certain tasks.
Others have tried designing virtual robot bodies and brains together, Bhatia says, but they have focused on simple tasks like walking and jumping. “One of the strengths of our work is the number of tasks and the number of unique tasks that we have developed,†he says.
Evolution Gym is open source: Bhatia’s team developed it to provide a benchmark platform on which every researcher can develop and test their own algorithms and compare approaches. In previous work, groups typically developed their own virtual environments for such assessments. The new digital space provides researchers with a common ground for measuring how well different algorithms stack up. “It enables people to measure progress – and that’s really important,†says Agrim Gupta, who has a doctorate in computer science. Student doing similar research at Stanford University. He recently published an article on how intelligence can not only be achieved through evolution, but can also be developed from experience. Bongard agrees, saying that the new MIT simulator will enable the field of evolutionary robotics to “move forward faster by clarifying which ways to develop robots work better than others”.
Such assessments are necessary because the robots designed by evolutionary algorithms do not always work. The MIT algorithms, for example, have not been able to successfully develop robots for catching and lifting. This shows that there is still a lot to be done in developing really intelligent robots, says Bhatia, which makes a standard platform like Evolution Gym even more important to jointly drive the development of robot design. “We enable the development of more intelligent AI algorithms in order to create real smart robots in the future,†he says.
Comments are closed.