Agile Flights Through a Moving Narrow Gap for Quadrotors Using Adaptive Curriculum Learning

Abstract

Fast and agile flying through a gap is challenging for a quadrotor if the gap is narrow, tilted, and moving. Due to the strict time-variant position and attitude constraints, collision-free traversal trajectories under the under-actuated quadrotor dynamics are sparse and difficult to solve. To achieve this challenging task in the real world, we propose a Gap-Traversing Adaptive Curriculum Learning (GTACL) approach, which consists of adaptive curriculum reinforcement learning (ACRL) and online thrust updating (OTU). First, ACRL is introduced to improve sample efficiency, and the policy training is accelerated by designing a curriculum adapted to the agent's capability. Second, OTU is proposed to map the acceleration commands to low-level throttle signals by estimating the thrust model during flight, which reduces the intermediate control variables and helps sim2real transfer. We use the prioritized experience replay mechanism that considers both policy update contribution and data acquisition time to adapt to the changing tasks. GTACL is trained entirely in simulation and can be transferred to other quadrotors with different dynamics. Furthermore, we achieve zero-shot transfer to the real-world quadrotor without fine-tuning. The average success rates of 98% and 87.8% in simulation and real-world experiments for different task conditions demonstrate the robustness of the proposed approach. Comparative results with traditional and related learning-based approaches show the advantages of GTACL in terms of learning efficiency, control performance, and generalization.