Learning-based Nonlinear Model Predictive Control to Improve Vision-based Mobile Robot Path Tracking
Defence Research and Development Canada Suffield, Alberta Canada
Pagination or Media Count:
This paper presents a Learning-based Nonlinear Model Predictive Control LB-NMPCalgorithm to achieve high-performance path tracking in challenging off-road terrain throughlearning. The LB-NMPC algorithm uses a simple a priori vehicle model and a learneddisturbance model. Disturbances are modelled as a Gaussian Process GP as a function ofsystem state, input, and other relevant variables. The GP is updated based on experiencecollected during previous trials. Localization for the controller is provided by an on-board,vision-based mapping and navigation system enabling operation in large-scale, GPS-deniedenvironments. The paper presents experimental results including over 3km of travel bythree significantly different robot platforms with masses ranging from 50 kg to 600 kg andat speeds ranging from 0.35 ms to 1.2 ms.1 Planned speeds are generated by a novelexperience-based speed scheduler that balances overall travel time, path-tracking errors,and localization reliability. The results show that the controller can start from a generica priori vehicle model and subsequently learn to reduce vehicle- and trajectory-specificpath-tracking errors based on experience.