Imagine a slime-like robot that could seamlessly change its shape to fit through tight spaces, which could unfold inside the human body to remove an unwanted item.
While such a robot does not yet exist outside of a laboratory, researchers are working to develop reconfigurable soft robots for applications in healthcare, wearable devices, and industrial systems.
But how can you control a soft robot that has no joints, limbs, or fingers that can be manipulated and can instead drastically alter its entire shape 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 that task requires the robot to change its morphology multiple times. The team also built a simulator to test control algorithms for deformable soft robots in a series of challenging shape-changing tasks.
Their method completed each of the eight tasks they tested and outperformed other algorithms. The technique worked especially well on multifaceted tasks. For example, in one test, the robot had to reduce its height while growing two tiny legs to fit through a narrow tube, and then grow those legs and extend its torso to open the tube's lid.
While reconfigurable soft robots are still in their infancy, such a technique could one day enable general-purpose robots that can adapt their shapes to perform various tasks.
“When people think of soft robots, they tend to think of robots that are elastic, but return to their original shape. Our robot is like a slime and can actually change its morphology. It is very surprising that our method worked so well because we are dealing with something very new,” says Boyuan Chen, a graduate student in electrical engineering and computer science (EECS) and co-author of a study. document on this approach.
Chen's co-authors include lead author Suning Huang, an undergraduate at Tsinghua University in China who completed this work while a visiting student at MIT; Huazhe Xu, assistant professor at Tsinghua University; and senior author Vincent Sitzmann, an assistant professor of EECS at MIT who leads the Scene Rendering Group in the Computer Science and artificial intelligence Laboratory. The research will be presented at the International Conference on Learning Representations.
Control 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 bring it closer to a goal.
This can be effective when the moving parts of the robot are consistent and well-defined, such as a gripper with three fingers. With a robotic gripper, a reinforcement learning algorithm could move a finger slightly, learning by trial and error whether that movement earns it a reward. Then I would move on to the next finger and so on.
But shape-shifting robots, controlled by magnetic fields, can dynamically flatten, bend or elongate their entire bodies.
Image: Courtesy of the researchers.
“Such a robot could have thousands of small pieces of muscle to control, so it is very difficult to learn it in a traditional way,” Chen says.
To solve this problem, he and his collaborators had to think about it differently. Instead of moving each small muscle individually, its reinforcement learning algorithm begins by learning to control adjacent muscle groups that work together.
Then, after the algorithm has explored the space of possible actions focusing on muscle groups, it drills down into finer details to optimize the policy or action plan it has learned. In this way, the control algorithm follows a coarse-to-fine methodology.
“Coarse to fine means that when you take a random action, that random action is likely to make a difference. The change in the result is probably very significant because several muscles are roughly controlled at the same time,” says Sitzmann.
To enable this, researchers treat a robot's action space, or how it can move in a given area, as an image.
Its machine learning model uses images of the robot's environment to generate a 2D action space, which includes the robot and the area around it. They simulate the robot's movement using what is known as the material point method, where the action space is covered by points, such as image pixels, and overlaid with a grid.
In the same way that nearby pixels in an image are related (like pixels that form a tree in a photo), they built their algorithm to understand that nearby action points have stronger correlations. The points around the robot's “shoulder” will move similarly when it changes shape, while the points on the robot's “leg” will also move similarly, but differently than those on the “shoulder.”
Additionally, researchers use the same machine learning model to observe 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 presents eight tasks that evaluate a reconfigurable robot's ability to dynamically change shape. In one, the robot must lengthen and curve its body in order to avoid obstacles and reach a target point. In another, it must change its shape to imitate the letters of the alphabet.
Image: Courtesy of the researchers.
“Our task selection in DittoGym follows both generic reinforcement learning reference design principles and the specific needs of reconfigurable robots. “Each task is designed to represent certain properties that we consider important, such as the ability to navigate through long-horizon scans, the ability to analyze the environment and interact with external objects,” says Huang. “We believe that together they can provide users with a comprehensive understanding of the flexibility of reconfigurable robots and the effectiveness of our reinforcement learning scheme.”
Their algorithm outperformed basic methods and was the only technique suitable for completing multi-stage tasks that required multiple shape changes.
“We have a stronger correlation between action points that are closer to each other, and I think that's key to making this work so well,” Chen says.
While it may be many years before shape-shifting robots are deployed in the real world, Chen and his collaborators hope their work will inspire other scientists not only to study reconfigurable soft robots but also to think about taking advantage of space. 2D action for other complex control problems.
