Learning Complex Motion Plans using Neural ODEs with Safety and Stability Guarantees

Learning from Demonstrations (LfD) is a framework that enables transfer of skills to robots from observations of desired tasks (Khansari-Zadeh and Billard, 2011; Ijspeert et al., 2013; Yang et al., 2022). Typically, the observations are robot trajectories that are demonstrated through kinesthetic teaching, passively guiding the robot through the nominal motion to avoid the correspondence problem (Akgun and Subramanian, 2011). In such a framework, it is essential to learn motion plans from as few demonstrations as possible, while still providing required robustness, safety, and reactivity in a dynamic environment. While there are multiple approaches to represent the motion, we focus on Dynamical Systems (DS) based formulation (Billard et al., 2022a). DS based approaches have been shown to be particularly useful in Human-Robot Interaction (HRI) scenarios (Khansari-Zadeh and Billard, 2011; Figueroa and Billard, 2022; Khansari-Zadeh and Billard, 2014), where the robot inherently adapts to changes in the environment and can be compliant to human interactions, instead of following a stiff time-dependent reference motion that encodes the task.