Sparse Control for Dynamic Movement Primitives

Patrick M. Wensing, Jean-Jacques E. Slotine

Abstract

Dynamic Movement Primitives (DMPs) are essential for generating complex robotic motions, yet they often lack formal stability guarantees in coupled or high-dimensional settings. This paper introduces a rigorous mathematical framework based on contraction theory to analyze DMP stability. We propose Sparsely-Inhibited Rhythmic DMPs (SI-RDMPs), demonstrating that a single, low-dimensional inhibitory signal can globally control high-dimensional rhythmic behaviors. Our approach is validated through numerical simulations and hardware experiments on the MIT Cheetah robot, showcasing smooth transitions and robust synchronization.

Introduction

In modern robotics, generating smooth, human-like trajectories often relies on Dynamic Movement Primitives (DMPs) and Central Pattern Generators (CPGs). While effective, these frameworks frequently operate without formal stability proofs, particularly when multiple primitives are coupled or scaled. This research addresses the critical need for modular stability guarantees and efficient control mechanisms for high-dimensional robotic systems.

Modular Stability via Contraction Theory

The core of this work utilizes contraction theory, a mathematical framework where stability is defined by the exponential convergence of trajectories toward one another. Unlike traditional Lyapunov stability, contraction analysis focuses on the "forgetting" of initial conditions, making it highly suitable for modular systems.

The authors extend this analysis to the standard DMP hierarchy, which typically consists of a reference system, a canonical system, and a transformation system.

Hierarchical structure of Discrete and Rhythmic DMPs showing contraction properties.

For rhythmic movements, which do not settle at a single equilibrium point, the paper employs transverse contraction analysis. This ensures that the system converges to a unique, stable limit cycle. Furthermore, the researchers demonstrate that DMPs can be scaled spatially and temporally through diffeomorphic transformations while preserving these fundamental stability properties.

Sparsely-Inhibited Rhythmic DMPs (SI-RDMPs)

A significant challenge in robotics is managing transitions—such as starting or stopping a complex gait—without manually adjusting every degree of freedom. The authors introduce SI-RDMPs to solve this through "sparse control."

By leveraging the properties of coupled contracting systems, they prove that a strong inhibitory signal applied to just a single node can stabilize an entire network. This stability "percolates" through the connections, forcing high-dimensional oscillations to collapse into a stationary state.

Network topology for sparse control where a single reference node (R) influences coupled canonical (C) and transformation (T) systems.

This mechanism mimics biological motor control, where simple neural triggers can activate or suppress complex, coordinated muscle patterns across the body.

Experimental Validation: The MIT Cheetah

The proposed framework was validated using both numerical simulations of heterogeneous oscillators and physical experiments on the MIT Cheetah quadruped.

In the hardware tests, the Cheetah was tasked with a "four-beat amble" gait. The SI-DMP framework enabled seamless start-stop transitions. For instance, when a sparse inhibitory signal was activated at $t = 108.3s$, the robot's leg oscillations converged to a halt. Upon removing the signal at $t = 112.4s$, the robot immediately and smoothly resumed its rhythmic gait. These results highlight the system's ability to achieve synchronization even when individual oscillators possess different natural frequencies.

Conclusion and Future Directions

This research provides a "modular stability toolkit" for roboticists, offering formal proofs for the widely used DMP architecture. While the current work focuses on internal dynamics, future iterations aim to incorporate external ground reaction forces and sensorimotor feedback to enhance robustness on uneven terrain. The concept of "decentralized inhibition" also presents an exciting avenue for more resilient, multi-limb coordination in unpredictable environments.

Summary updated

BibTeX
@inproceedings{wensing2017sparse,
  author = {Patrick M. Wensing and Jean-Jacques E. Slotine},
  title = {{Sparse Control for Dynamic Movement Primitives}},
  booktitle = {IFAC-PapersOnLine},
  year = {2017},
  doi = {10.1016/j.ifacol.2017.08.1789},
}
Figure