Magnetic Robot Position and Angle Transition Dataset for MLP Training

Magnetic soft robots, known for infinite degrees of freedom and fast actuation, pose significant control challenges for precision-sensitive biomedical applications. Current model-free control methods using reinforcement learning are robot-specific, requiring each individual robot's precise material and physical properties. We thus propose a novel generalized model-free control approach for magnetic soft robots, independent of the robot shape and material properties. We use on-policy reinforcement learning to autonomously create a representative dataset of state-action pairs, assuming the absence of expert control demonstrations and train multi-layer perceptron ensemble to predict successive spatial states. A* path-planning is further incorporated to generate input signal action sequences. This approach was validated with autonomous multi-target research navigation, and results demonstrated improved prediction accuracy for spatial states by 80\% as compared to a single MLP reference model. The proposed novel model free closed-loop control approach for soft robots could facilitate diverse applications of magnetic soft robots for minimally invasive surgery, bio-safe object manipulation and remote drug delivery.