Obstacle-aware Simultaneous Task and Energy Planning with Ordering Constraints
Di Wang · 2023
To improve the practical performance of task planning for unmanned ground vehicles, geographical features, limited onboard energy, and ordering constraints are urgent to be considered. This paper studies the simultaneous task and energy planning (STEP) problem considering obstacles and ordering constraints. The STEP problem is a sequential decision-making problem using the Markov decision process. A new multi-head self-attention-based deep reinforcement learning (DRL) method is proposed to solve this problem. A distance estimator calculates the distance between any two task points considered obstacles. A relational network pairs ordering constraints with feature vectors of each task to reason their relationships. The simulations compare solutions of our method, a recurrent neural network (RNN) based DRL, and the CPLEX method. Results demonstrate that our approach can obtain approximated performance compared with RNN-based DRL but in a shorter training time.