Towards Intelligent Soft Robots: A Data-Driven Approach to Unlock the Potential of Magnetically Responsive Materials

Soft robots have emerged as a new branch of robotics with deformable bodies to achieve adaptability to dynamically changing unstructured environments and safe interaction with life forms ranging from cells to humans. Miniaturization efforts in soft robotics for operation in confined, small, and remote spaces are challenging because conventional batteries, actuators, and on-board processors are not available at smaller scales. The challenges of power, intelligence, and actuation at smaller scales have led to the incorporation of stimuli-responsive materials in soft robotics, where the actuation mechanism is based on the encoded response of the robot body to the external stimuli at the material level. Among the wide range of proposed external stimuli such as temperature, light, chemical, electric and magnetic fields, magnetic fields are particularly promising due to their safe and transparent interaction around biological tissues. Magnetically responsive soft materials with fast response and untethered control capabilities at small scales in confined environments make them ideal candidates for minimally invasive clinical operations within the human body. Advances in the spatial programming of magnetic soft materials have enabled micron-scale high resolution, 3D directionality, and multi-material compositions, resulting in a vast design space with intricate coupled magnetic and mechanical responses of soft robots. While the design space is immensely enriched, existing designs rely on an intuitive trial-and-error approach with fixed morphologies and simplified magnetic programming, drastically limiting the achievable capabilities. Here, we present a generic and experience-free method for programming both the 3D structural design and the 3D magnetic profile of magnetic soft robots for desired 3D shape morphing and behavior. The design algorithm is based on a continuous exploration of the design space through design candidates generated by MAP elites and the exploitation of promising designs guided by a predictive neural network (NN) model. The selected promising designs are tested in a computationally inexpensive simulation engine capable of evaluating the dynamic behavior of magnetic soft robots. The resulting best-performing designs for desired 2D and 3D shape-morphing of beams inspired by mathematical functions and complicated sharp-cornered objects are experimentally demonstrated, demonstrating the Sim2Real transfer. The developed design strategy is also used to achieve desired dynamic behaviors without a defined target shape-morphing, including maximizing the number of turns, maximizing the height, and minimizing the bounding sphere volume for soft magnetic beams and plates. The superiority of the data-driven design strategy is highlighted in the high-performance jumping behavior of magnetic soft robots, where the intuitive design adapted from the literature failed to lift off the surface. Furthermore, the data-driven designed robots are shown to be capable of complex locomotion behaviors with simple control signals compared to the fine-tuned complex ones. The data-driven design strategy presented here provides a systematic and versatile platform to bridge the gap between advances in material functionalities and fabrication and programming methods, thus unlocking the potential of stimulus-responsive soft robots for real-world applications.