Non-Smooth Dynamics of Constrained Task-Oriented Dynamical Systems
Abstract
Major Goals: Power, physical stress, and force constraints are common in many human endeavors, but these constraints are also prevalent in robotics, autonomous vehicles, and other engineered systems. In addition to being constrained by power output, human and animal muscles are also constrained by a maximal force-elongation velocity curve. The fact that all humans and animals are limited by such constraints requires them to think differently to achieve many complex tasks. It is the goal of this proposal to constrain the system actuation in our experimental systems, with one or more constraints; multiple methods will then be investigated to enable the complex dynamical systems to learn to achieve specific tasks. Accomplishments: We have developed a model-free framework and forecasting strategy to enable dynamical systems to learn to efficiently exploit their natural dynamics. The work builds upon past literature in the area of reinforcement learning to advance the current understanding for nonlinear dynamical systems. We have also completed an initial investigation of constraining an actuator while building up momentum to achieve an attractor escape. The aforementioned ideas have been applied to the problem of nonlinear systems switching attractors. The focus has been on constrained actuation and limiting the energy expenditure when applying control.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 30, 2021
- Accession Number
- AD1196551
Entities
People
- Brian P. Mann
Organizations
- Duke University