Probabilistic Movement Primitives
Overview
Movement primitives (MPs) are a well-established approach for representing interchangeable and reusable generators of robot movement. They provide a condense parameterization of a robot's control policy. In robot learning, multiple MPs may be combined as building blocks in a modular control architecture to solve complex tasks. They can be blended between motions, adapted to altered task variables, and co-activated in parallel. Many robot learning successes are based on MPs due to their ability to accurately approximate continuous and high-dimensional movements. For example, MPs have been employed in learning from demonstration using imitation learning. Probabilistic movement primitives (ProMPs) are a concept whereby a distribution of trajectories is learned from multiple demonstrations. They provide an effective approach to learning flexible trajectories from demonstration.
Contributions
- We designed a ProMP controller that guarantees the system never leaves a neighborhood defined by the training set, and it provides a simple way to define trajectories that enforce safety constraints for manipulation tasks
- We combined model predictive control and control barrier functions to guide a robot along a predefined trajectory while guaranteeing it always maintains a desired distance from a human motion distribution defined by a ProMP
- We developed a novel means of automating the design of CLFs and CBFs from the distribution delivered by a ProMP and demonstrated its practical applicability through experimental validation on a Universal Robots UR5e
- We created a safe control design approach that takes demonstrations provided by a human teacher to enable a robot to accomplish complex manipulation scenarios in dynamic environments
Publications
-
M. Davoodi, A. Iqbal, J.M. Cloud, W.J. Beksi, and N.R. Gans,
"Probabilistic Movement Primitive Control via Control Barrier Functions,"
International Conference on Automation Science and Engineering (CASE), 2021.
Best Conference Paper Award Finalist • Paper • Citation -
M. Davoodi, J.M. Cloud, A. Iqbal, W.J. Beksi, and N.R. Gans,
"Safe Human-Robot Coetaneousness Through Model Predictive Control Barrier Functions and Motion Distributions,"
Modeling, Estimation, and Control Conference (MECC), 2021.
Paper • Citation -
M. Davoodi, A. Iqbal, J.M. Cloud, W.J. Beksi, and N.R. Gans,
"Safe Robot Trajectory Control using Probabilistic Movement Primitives and Control Barrier Functions,"
Frontiers in Robotics and AI, 2022.
Paper • Citation -
M. Davoodi, A. Iqbal, J.M. Cloud, W.J. Beksi, and N.R. Gans,
"Rule-Based Safe Probabilistic Movement Primitive Control via Control Barrier Functions,"
IEEE Transactions on Automation Science and Engineering, 2022.
Paper • Citation