Imagine a slime-like robot that can change its shape to seamlessly squeeze through tight spaces, which could be deployed to remove an unwanted object inside the human body.
Although such a robot does not yet exist outside of a laboratory, researchers are working to develop programmable soft robots for applications in healthcare, wearable devices and industrial systems.
But how does one control a squishy robot that has no joints, limbs, or fingers to manipulate, and instead can completely change its entire appearance at will? MIT researchers are working to answer that question.
They developed a control algorithm that can autonomously learn how to move, stretch, and shape a reconfigurable robot to complete a specific task, even when the robot for that task Needs to change its shape multiple times. The team also built a simulator to test control algorithms for impractical soft robots on a series of challenging, shape-shifting tasks.
Their method performed better than other algorithms on each of the eight tasks they evaluated. The technique worked particularly well on multidimensional tasks. For example, in one test, the robot had to lower its height while extending two short legs to squeeze through a narrow pipe, and then extend those legs and raise its torso to open the pipe lid. .
Although reconfigurable soft robots are still in their infancy, such techniques may someday enable general-purpose robots that can adapt their shapes to accomplish different tasks.
“When people think of soft robots, they think of robots that are flexible, but spring back to their original shape. Our robot is like slime and can actually change shape. It It's surprising that our method works so well because we're dealing with something so new,” says Boiyuan Chen, a graduate student in Electrical Engineering and Computer Science (EECS) and co-author of a paper on the approach. .
Chen's co-authors include lead author Xing Huang, an undergraduate student at Tsinghua University in China who completed the work while a visiting student at MIT. Huazhe Xu, an assistant professor at Tsinghua University; and senior author Vincent Seitzman, assistant professor of EECS at MIT who leads the Scene Representation Group in the Computer Science and Artificial Intelligence Laboratory. This research will be presented at the International Conference on Learning Representations.
Controlling dynamic movement
Scientists often teach robots to complete tasks using a machine learning approach known as reinforcement learning, which is a trial-and-error process in which the robot is rewarded for actions that make it a Get closer to the goal.
This can be useful when the moving parts of the robot are continuous and well defined, such as a gripper with three fingers. With a robotic gripper, a reinforcement learning algorithm can move a finger slightly, learning by trial and error whether that movement brings a reward. Then it moves to the next finger, and so on.
But shape-shifting robots, controlled by magnetic fields, can dynamically squish, bend or lengthen their entire bodies.
“A robot like this can have thousands of tiny muscle segments, so it's very difficult to learn in a traditional way,” Chen says.
To solve this problem, he and his colleagues had to think about it differently. Instead of moving each small muscle individually, their reinforcement learning algorithm begins by learning to control adjacent muscle groups working together.
Then, after the algorithm explores the space of possible actions by focusing on muscle groups, it learns to fine-tune the policy, or action plan, it exercises. Thus, the control algorithm follows a coarse-to-fine procedure.
“Coarse-to-fine means that when you do a random action, that random action is likely to make a difference. The potential for variation in results is significant because you're coarsening many muscles at the same time. control,” says Seitzman.
To enable this, researchers treat the robot's action space, or how it can move in a specific area, such as a picture.
Their machine learning model uses images of the robot's environment to create a 2D action space, which includes the robot and its surroundings. They simulate robot motion using what is called the material point method, where the action space is covered with points, such as image pixels, and overlayed with a grid.
As nearby pixels in an image are interconnected (like the pixels that make up a tree in an image), they built their algorithm to understand that nearby action points are strongly correlated. The points around the robot's “shoulder” will move as it changes shape, while the points on the robot's “leg” will also move, but in a different way than the “shoulder”.
In addition, the researchers use the same machine learning model to look at the environment and predict the actions the robot should take, making it more efficient.
Building a simulator
After developing this approach, the researchers needed a way to test it, so they created a simulation environment called DittoGym.
DittoGym features eight tasks that test a reconfigurable robot's ability to dynamically change shape. In one, the robot must lengthen and curve its body to weave around obstacles to reach a target location. In another, it must change its shape to mimic the letters of the alphabet.
“Our task selection at DittoGym follows both general reinforcement learning benchmark design principles and the specific needs of configurable robots. Each task is designed to represent certain characteristics that we consider important, such as the ability to navigate through long-horizon searches, analyze the environment, and interact with external objects,” says Huang. “We believe that together they can give users a comprehensive understanding of the flexibility of reconfigurable robots and the effectiveness of our reinforcement learning scheme.”
Their algorithm outperformed baseline methods and was the only technique suitable for completing multistage tasks that required many shape transformations.
“We have a strong connection between action points that are close together, and I think that's the key to making this work so well,” says Chen.
Although it may be years before shape-shifting robots are deployed in the real world, Chen and his colleagues hope their work will inspire other scientists not only to study programmable soft robots, but also for other complex control problems. 2D action also encourages thinking about exploiting space.