Researchers at MIT are working on developing reconfigurable soft robots that can change their shape to complete specific tasks. These robots have the potential to be used in various applications, including health care, wearable devices, and industrial systems. A control algorithm has been developed that allows these robots to autonomously learn how to move, stretch, and shape themselves to complete tasks. The team has also built a simulator to test control algorithms for deformable soft robots on challenging, shape-changing tasks, with their method outperforming other algorithms.
The research team, led by MIT professor Vincent Sitzmann, includes graduate student Boyuan Chen and undergraduate student Suning Huang from Tsinghua University. The algorithm uses a machine-learning approach known as reinforcement learning, where the robot is rewarded for actions that move it closer to the goal. Instead of moving each individual muscle of the robot, the algorithm first learns to control groups of adjacent muscles that work together before optimizing the action plan. This coarse-to-fine methodology allows the robot to dynamically squish, bend, or elongate its entire body using a magnetic field.
The researchers treat a robot’s action space as an image, using the material-point-method to simulate robot motion with points overlayed on a grid. The algorithm is designed to understand that nearby action points have strong correlations, similar to nearby pixels in an image. This approach enables the robot to adapt its shape based on its environment and predict the actions it should take more efficiently. The team created a simulation environment called DittoGym to test their approach, with tasks that evaluate the robot’s ability to dynamically change shape, outperforming baseline methods and completing multistage tasks that require multiple shape changes.
While it may be some time before shape-shifting robots are deployed in real-world applications, the researchers hope that their work will inspire others to explore reconfigurable soft robots and leverage 2D action spaces for complex control problems. The ability to control deformable robots that can change shape to accomplish tasks opens up new possibilities for robotics in various fields. The researchers believe that their method of using reinforcement learning to control shape-shifting robots has the potential to revolutionize the field of soft robotics and lead to the development of general-purpose robots that can adapt their shapes to a wide range of tasks.
