Global Reorientation of a Free-Fall Multibody System Using Reconstruction Loss-Based Deep Learning Method

Inspired by animal’s righting flex behavior, a free-fall robotic system with zero angular momentum can orient one or more its bodies by periodically moving some of its joints. However, considering the complex nonlinear dynamics and nonholonomic constrains, planning such periodical closed-loop joint motions for achieving a required orientation of one specific body (such as the head or base of the robot) remains challenging in solution’s accuracy, simplicity and generalization. In this paper, we propose a deep learning-based method to address the challenge. First, the method employs a reconstruction loss for training with non-gradient simulation data, which increases the solution’s accuracy. Second, it approximates general solutions of the system and expands the reachable rotation in a single motion to reduce the repetition of periodical motions. Third, with the generalization of neural net-works, the method can be applied to systems with unknown physical proper-ties and can plan the needed periodical motions in real time after training. As an example, the method has been implemented on a squirrel-like quadruped model and its simulation results are compared with other baseline reinforcement learning and inverse kinematic methods.