QPRL: Learning Optimal Policies with Quasi-Potential Functions for Asymmetric Traversal

Overview

In goal-reaching tasks, a transition can be easy to make and difficult to undo. QPRL studies a representation of traversal cost that explicitly accounts for that directional structure.

The method decomposes cost into a potential difference and a path-dependent residual. This structure is combined with a Lyapunov-based constraint during policy learning.