Learning Stabilizable Dynamical Systems via Control Contraction Metrics
share
We propose a novel framework for learning stabilizable nonlinear dynamical systems for continuous control tasks in robotics. The key idea is to develop a new control-theoretic regularizer for dynamics fitting rooted in the notion of stabilizability, which guarantees that the learnt system can be accompanied by a robust controller capable of stabilizing any trajectory that the system can generate. By leveraging tools from contraction theory, statistical learning, and convex optimization, we provide a general and tractable algorithm to learn stabilizable dynamics, which can be applied to complex underactuated systems. We validate the proposed algorithm on a simulated planar quadrotor system and observe that the control-theoretic regularized dynamics model is able to consistently generate and accurately track reference trajectories whereas the model learnt using standard regression techniques, e.g., ridge-regression (RR) does extremely poorly on both tasks. Furthermore, in aggressive flight regimes with high velocity and bank angle, the tracking controller fails to stabilize the trajectory generated by the ridge-regularized model whereas no instabilities were observed using the control-theoretic learned model, even with a small number of demonstration examples. The results presented illustrate the need to infuse standard model-based reinforcement learning algorithms with concepts drawn from nonlinear control theory for improved reliability.
READ FULL TEXT