If I asked you to create the perfect robot for throwing blocks, you would probably suggest something humanoid with hands and legs. Who could blame you? Throwing stuff is what humans excel at.
The odds are that you thought instead of the robot in the video, which is a Frankenstein’s monster made up of pieces of Tetris. This is because a computer “evolved” the robot’s body and brain in a new platform from MIT’s Computer Science and Artificial Intelligence Laboratory. The Evolution Gym is where robots of the future can trust algorithms to design their designs instead of relying upon human designers (and their biases). Wojciech Matusik from MIT, a co-creator, says that the system has the potential to discover new robot designs and also has the potential to produce more high-performing robots. How much intelligence can be created if you start with very, very basic structures?
It turns out, a lot. You can see the soft robot learning how to perform flips. Each “voxel” or unit within this robot is color-coded. Gray is a piece that does not do any active work, while black signifies a rigid piece. The colored voxels, on the other hand, are actuators orbits of any robot that can produce movement. These are the motors that drive rigid robots. The actuator expands or contracts vertically in blue, and contracts horizontally in orange. The robot is basically turning on its black rigid voxel, and the colored actuators are pushing it off the ground.
This robot, however, is simple and consists of only horizontally-actuating voxels. It’s legs, basically. The voxels allow the machine to move with astonishing grace, a kind of gallop, by expanding and contracting in coordination.
Here’s one who can climb. The blue actuators are located on either side. They alternate their movements to grab the surface. A sort of appendage on the top of this soft robot guides it up the column. The machine must adapt to the soft sections of the column as it moves up. This is a complex behavior for a robot that was so simple, and even more difficult than one that was designed.
In order to create hundreds upon hundreds of morphologies, the system starts by combing those four types of voxels together in unique ways. This is similar to how biological evolution created the different forms of humans (legs), animals (no legs), or birds (wings). Each column in the above image shows the results of four generations. These robots are also optimized by the algorithm, which basically optimizes the brains of the actuating voices to perform their tasks in the right order.
You can have a good time in the Evolution Gym. The different morphologies, each with a different brain, are scored according to how they do a task such as walking forward or throwing blocks. Matusik says, “You choose the soft robots with the highest performance, and you’ll evolve them or modify them until you refine their structures.” It’s survival by the fittest: Robots that can walk and climb the most are the best to seed the next generation. The rest of us get tossed out. There are evolutionary robots. Similar techniques are being used by other researchers to optimize certain surfaces for robot legs.
This video shows the evolution of a block-carrying robotic, which eventually becomes a galloping machine capable of balancing the object on its back.
These designs can adapt to the environment and even overcome obstacles, as this robot that is irregularly shaped shows by bounding across an uneven surface.
Here’s another one that can be used to manipulate an overhead beam.
These robots were created by algorithms using simulations, and don’t exist. These designs are very limited at the moment. A particular morphology may be able to climb or throw a block, but not both. The Evolution Gym, however, is open-source and can be used by any researcher to refine these techniques.
Matusik says, “This environment can potentially stimulate much new research.” “You can create intelligent robotic systems that are more efficient and flexible from scratch without having to have any preconceived ideas about the brain or structure.” After designing and training machines in the gym, the next step would be to build the best candidates out in the real world. This will ensure that researchers don’t waste their time creating prototypes only to discover it doesn’t work.
Tonnes Nygaard studies evolutionary robotics at The Norwegian Defence Research Establishment but was not involved in this project. “I am very excited about seeing how other researchers begin using this benchmarking framework and what kind of creative virtual robots emerge from it in the end,” Tonnes Nygaard said. “Any system that fosters interest and collaboration in research like this one is welcomed with open arms!”
