Physics-informed data-driven control for energy-efficient spherical robot maneuvering

Spherical robots operating in unstructured environments face substantial challenges due to their nonlinear, underactuated dynamics and onboard power limitations. Traditional model-based controllers often lose effectiveness under dynamic uncertainties and actuator constraints, while purely data-driven schemes may require extensive training or produce unrealistic control signals. This paper proposes a hybrid, physics-informed, data-driven control strategy for a pendulum-driven spherical robot. A neural network (NN) controller is trained in simulation only with a loss function that penalizes trajectory errors, large control signals, and torque saturation. The control law leverages a modified terminal sliding mode formulation that guarantees finite-time convergence, validated through Lyapunov arguments. Numerical simulations demonstrate that the proposed method consumes significantly less control energy than conventional strategies and remains robust under disturbances, thus promoting energy-efficient trajectory tracking and offering a promising solution to the challenges faced by spherical robots in unstructured environments.