Energy Aware Reinforcement Learning for Control of Autonomous Surface Vehicles Subjected to Uncertai
Abstract
The safe and robust operation of autonomous surface vehicles (ASVs) has become a critical challenge when they are deployed and opera,ted in uncertain dynamic environments, where there is the participation of additional agents such as other vessels. The movement of,external agents translates into a safety problem in practice that is very important when defining the routes of ASVs. The precise pr,ediction of the movement of an agent is a very difficult task to carry out due to incomplete knowledge and, therefore, to the uncert,ainty with which these movements are subjectively conceived. On the other hand, every time evasive maneuvers have to be executed to,avoid colliding, time and energy are spent as the vehicle deviate from the minimum energy path (defined in path-planning) and actuat,ors might even require higher energy consumption to maneuver sharply. Although collision avoidance is a critical aspect of autonomou,s navigation, energy autonomy is important as well, and its neglect could eventually lead to a catastrophic failure in which the veh,icle is lost or causes an accident.The objective of this project is to formulate energy-aware autonomous navigation of ASVs under a,probabilistic approach of Chance-Constrained Nonlinear Model Predictive Control (CCNMPC) and reinforcement learning (RL) and to deve,lop algorithms that allow solving it in real-time so that a formal procedure can be established to automatically determine the most,appropriate risk aversion to collision conditional on the risk of loss of energy resources. This establishes an energy criterion to,restrict maneuverability and risk aversion of collision in favor of guaranteeing energy autonomy for a safe arrival at the destinati,on or mission replanning.The expected results of this project include the development and experimental validation of new control str,ategies based on CCNMPC for the navigation of ASVs in dynamic environments that conceive in the form of nested loops the problem of,collision avoidanceand the problem of energy autonomy; all within a probabilistic framework. With this, it is expected to lower the,computational cost of the collision avoidance scheme by penalizing energy consumption in the form of simple restrictions to compute,, which will be defined from an outer control loop whose dynamics are slower than the control to avoid collisions, and whose purpose,is to appropriately distribute the energy throughout the navigation based on statistical learning about the dynamic environment.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 06, 2022
- Source ID
- N629092212056
Entities
People
- Giancarlo Troni
Organizations
- Office of Naval Research
- Pontifical Catholic University of Chile
- United States Navy