Scientists at MIT have been working on the design and control of a reconfigurable, squishy, soft robot, similar in nature to ‘slime’, that has potential applications in healthcare, wearable devices and industrial systems due to its ability to shape-shift to complete varying tasks. These soft robots currently only exist in labs and do not possess defined features such as joints, fingers, or limbs that can be moved. They can alter their shape as needed with a control algorithm developed by the MIT researchers. This algorithm autonomously learns how to instruct the robot to move, stretch and alter its form to accomplish a task.
The research team developed a simulator called DittoGym to examine the control algorithms. The robot was given multifaceted tasks, like reducing its height while growing small legs to pass through a narrow pipe and then retract the legs and extend its torso to open a pipe’s lid. The robot successfully completed each task, showcasing the possibility of future robots being capable of adapting their shapes to various tasks.
The control system utilized reinforcement learning, a machine learning approach whereby the algorithm is rewarded for performing tasks that help achieve a given goal. Standard robots may use this system to manage their actions, such as a gripper closing or opening. However, a fluid robot like the one developed by the MIT researchers required a different approach. Rather than individual movements, their algorithm commences by learning to manage groupings of muscles. The algorithm narrows down the optimized plan of action after it explores potential actions.
Successful control of the robot was achieved using a 2D action space that simulates its environment, which helped predict effective movements for the task at hand. The simulator, DittoGym, was used to assess its ability to dynamically change shape, with the robot being asked to elongate and curve to move around obstacles, or to change form to resemble letters of the alphabet.
MIT’s research results appear promising, with the algorithm outperforming other methods and being the only technique that could perform multi-stage tasks involving several shape changes. However, scientists predict it will be many years before such shape-shifters can be effectively used outside the lab. The researchers behind the project hope that their work will inspire others to study reconfigurable soft robots further or to consider using the 2D action space model for other complex control problems.
