Efficient Sampling-based GP-MPC for autonomous robots

Most control methods operate under the assumption of a known model. However, in practice, knowing the exact dynamics model a priori is unrealistic. A common approach is to model the unknown dynamics using Gaussian Processes (GPs) which can characterize uncertainty and formulate a Model Predictive Control (MPC) type problem. However, it is difficult to exactly utilize this uncertainty characterization in predictive control. In a recent approach [1], we proposed a sampling-based robust GP-MPC formulation for accurate uncertainty propagation by sampling continuous functions. In contrast, in the proposed project, you will implement an approximation method for sampling continuous functions using a finite number of basis functions [2] and solve the MPC problem jointly with the sampled dynamics. You will analyze the trade-offs between performance, approximation accuracy, and computational cost for this method.

Keywords: Gaussian Processes, model predictive control (MPC), Issac sim simulator, uncertainty propagation, Algorithms

Calendar

Location

Research Zeilinger (ETHZ)

Labels

Topics

• Engineering and Technology

Documents

Name	Comment	Size	Actions
Efficient sampling based GPMPC for autonomous robots (1).pdf		162KB	Download