AgilePilot: DRL-Based Drone Agent for Real-Time Motion Planning in Dynamic Environments by Leveraging Object Detection

TL;DR

The paper tackles real-time motion planning for drones in dynamic environments with moving gates and obstacles. It introduces AgilePilot, a DRL-based planner trained in dynamic simulation (Gym PyBullet) and augmented with a real-time computer vision stack (YOLOv8n Pose, IPPE PnP, EKF) to estimate poses and predict velocities. A reward-structured, Sim2Real training framework yields safety-aware agility and achieves up to $3.0$ m/s in real-world tests, outperforming an APF-based planner by about 3x in tracking accuracy and speed, with a $90\%$ success rate across 75 trials. Hardware experiments validate real-time perception and dynamic navigation, showing modest pose-estimation errors (RMSE ≈ $0.076$ m for obstacles and ≈ $0.37$ m for gates) and robust operation in challenging conditions, supporting practical deployment of agile drones.

Abstract

Autonomous drone navigation in dynamic environments remains a critical challenge, especially when dealing with unpredictable scenarios including fast-moving objects with rapidly changing goal positions. While traditional planners and classical optimisation methods have been extensively used to address this dynamic problem, they often face real-time, unpredictable changes that ultimately leads to sub-optimal performance in terms of adaptiveness and real-time decision making. In this work, we propose a novel motion planner, AgilePilot, based on Deep Reinforcement Learning (DRL) that is trained in dynamic conditions, coupled with real-time Computer Vision (CV) for object detections during flight. The training-to-deployment framework bridges the Sim2Real gap, leveraging sophisticated reward structures that promotes both safety and agility depending upon environment conditions. The system can rapidly adapt to changing environments, while achieving a maximum speed of 3.0 m/s in real-world scenarios. In comparison, our approach outperforms classical algorithms such as Artificial Potential Field (APF) based motion planner by 3 times, both in performance and tracking accuracy of dynamic targets by using velocity predictions while exhibiting 90% success rate in 75 conducted experiments. This work highlights the effectiveness of DRL in tackling real-time dynamic navigation challenges, offering intelligent safety and agility.

Figures (13)

• Figure 1: AgilePilot technology performs deep reinforcement learning-based motion planning to avoid obstacles and navigate through moving gates.
• Figure 2: AgilePilot consists of a position estimation module for the agent and dynamic objects in the environment, along with a control policy that maps state observations to control commands.
• Figure 3: Gym PyBullet custom simulation environment.
• Figure 4: Neural network architecture representing the input layer, hidden layers, and output layer of our system.
• Figure 5: Colormap visualizes reward values for drone positions in the presence of an obstacle and a goal in a defined area, i.e., (a) shows reward structure in 3D space where the obstacle is considered as infinite height, and (b) shows a top view in the XY plane with the predicted path of the drone that avoids the obstacle and reaches the goal.