This paper develops a navigation algorithm for vehicles in complex environments, combining a hard guarantee of constraint satisfaction with the ability to learn from successive missions. The method uses receding horizon control for constrained short-term path planning and control, with the cost-to-go developed by reinforcement learning. Simulation results show that the algorithm learns to reproduce the shortest path through an environment with obstacles and can determine good behaviours for multiple surveillance tasks.
|Translated title of the contribution||Combining Planning and Learning for Autonomous Vehicle Navigation|
|Title of host publication||AIAA Guidance Navigation and Control Conference, Toronto|
|Publication status||Published - Aug 2010|
Bibliographical noteConference Organiser: AIAA
Other identifier: AIAA-2010-7866