Stable Machine Learning for Prediction and Control
This project focuses on developing physics-contrained machine learning techniques that improve both short- and long-term prediction accuracy. We develop numerical methods that guarantee the stability of data-driven modeling, prediction, and control and can run in real time. We test these methods in simulations and experiments on various robotic systems and applications, including predicting the gait cycle of a biped, stabilizing a quadrotor, and tracking trajectories or pushing blocks using the Franka Emika robot.
Experiments
To illustrate the importance of stabiity on data-driven control, I use the Franka Emika robot to apply control based on a least-squares, unstable data-driven model and my stability-enforcing data-driven algorithm.
As seen from the simulation, the unconstrained least-squares solution is not safe to test on the physical robot.
The performance of the constrained, stable data-driven model is shown next, in simulation and experiment.
Pusher-Slider
I also test the algorithm on a pusher-slider system that has hybrid dynamics; the Franka Emika manipulator holding a stick pushes a block to a target.Before the experiments, I measure the accuracy of the stable data-driven model in simulation. I compare the real trajectory of the measured training data to the prediction of the model, given the applied controls. The stable data-driven model accurately captures the hybrid dynamics between the manipulator and the block and successfully pushes it to the target when tested in experiments.
Publications
Data-Driven Identification of Stable Koopman Operators
G. Mamakoukas and T. D. Murphey
Science Robotics, In Preparation , 2020.
Learning Stable Linear Systems for Prediction and Control
G. Mamakoukas and T. D. Murphey
Double blind peer reviewed venue , Submitted, 2020.