Data Driven Controller Design

We have proposed a generalized control policy for rapid learning and control of unknown dynamics such as human-robot interaction or unmodeled disturbance, using Gaussian Process Regression (GPR). GPR provides granular control for the tradeoff between bias and variance which is very critical in hyperparameter tuning and also influence the speed of training. Here GPR is used to learn the generalized global controller using the local trajectories as the training data. These local trajectories are optimized using an iterative linear quadratic regulator (iLQR) and differential dynamic programming (DDP). Dynamics used in this trajectory optimization is estimated from the sensor data using the Gaussian mixture model (GMM) prior modeling. This generalization was tested on the cart-pole problem and the results illustrate for singular rollout using the generalized global policy was able to predict 91% of the control input when compared to best possible policy, using known dynamics. Inferring from these results, we suggest that GPR could be an alternative to the present global policy optimizers such as neural network, solidifying the usage in environment independent control of interactive controllers.