A Dynamically Feasible Fast Replanning Strategy with Deep Reinforcement Learning

Hasanzade M. , Koyuncu E.

JOURNAL OF INTELLIGENT & ROBOTIC SYSTEMS, cilt.101, sa.1, 2021 (SCI İndekslerine Giren Dergi) identifier identifier

  • Cilt numarası: 101 Konu: 1
  • Basım Tarihi: 2021
  • Doi Numarası: 10.1007/s10846-020-01274-1


In this work, we aim to develop a fast trajectory replanning methodology enabling highly agile aerial vehicles to navigate in cluttered environments. By focusing on reducing complexity and accelerating the replanning problem under strict dynamical constraints, we employ the b-spline theory with local support property for defining the high dimensional agile flight trajectories. We utilize the differential flatness model of an aerial vehicle, allowing us to directly map the desired output trajectory into input states to track a high dimensional trajectory. Dynamically feasible replanning problem is addressed through regenerating the local b-splines with control point reallocation. As the geometric form of the trajectory based on the location of the control points and the knot intervals, the control point reallocation for fast replanning with dynamical constraints is turned into a constrained optimization problem and solved through deep reinforcement learning. The proposed methodology enables generating dynamically feasible local trajectory segments, which are continuous to the existing, hence provides fast local replanning for collision avoidance. The DRL agent is trained with different environmental complexities, and through the batch simulations, it is shown that the proposed methodology allows to solve fast trajectory replanning problem under given or hard dynamical constraints and provide real-time applicability for such collision avoidance applications in agile unmanned aerial vehicles. Hardware implementation tests of the algorithm with the agile trajectory tracker to a small UAV can bee seen in the following video link: https://youtu.be/8IiLQFQ3V0E.