Trajectory Planning With Deep Reinforcement Learning in High-Level Action Spaces

This article presents a technique for trajectory planning based on parameterized high-level actions. These actions are subtrajectories that have variable shape and duration. The use of can improve the performance guidance algorithms. Specifically, we show how improves policies generated via reinforcement learning (RL). RL has shown great promise solving complex control, guidance, coordination problems but still suffer from long training times poor performance. work shows reduces required number steps increases path an RL-trained policy. We demonstrate method space-shuttle example. proposed (latitude range) by 18% compared with baseline implementation. Similarly, achieves steady state during approximately 75% fewer steps. also policy enables effective in obstacle field. Finally, this develops loss function term policy-gradient-based deep RL, which is analogous to antiwindup mechanism feedback control. inclusion underlying optimization average return our numerical