AgilePilot: DRL-Based Drone Agent for Real-Time Motion Planning in Dynamic Environments by Leveraging Object Detection (2502.06725v2)

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.

• The paper introduces a DRL-based motion planning algorithm that integrates real-time object detection for agile drone navigation.
• It employs a PPO-based actor-critic model within a Gym PyBullet simulation to train drones for dynamic obstacle avoidance.
• Experiments demonstrate a 90% success rate and tripled performance compared to traditional APF methods in dynamic settings.

AgilePilot: A DRL-Based Approach to Drone Navigation in Dynamic Environments

The paper "AgilePilot: DRL-Based Drone Agent for Real-Time Motion Planning in Dynamic Environments by Leveraging Object Detection" presents a novel approach in the domain of autonomous drone navigation using deep reinforcement learning (DRL). This research addresses the fundamental challenges faced by drones operating in dynamic and unpredictable environments, where traditional planners often struggle due to their limitations in real-time adaptability and decision-making.

Summary of Methodology

The core contribution of this paper is the development of AgilePilot, a motion planning algorithm that leverages DRL to enable drones to navigate environments with rapidly changing conditions. The proposed system incorporates real-time computer vision (CV) for object detection, allowing drones to adjust their trajectories dynamically based on the movement of targets and obstacles. The training-to-deployment framework mitigates the Sim2Real gap, crucial for effective real-world implementation. The DRL model is trained using a sophisticated reward structure that balances safety and agility according to environmental conditions.

This paper uses a Gym PyBullet simulation environment for training that includes custom models for drones and dynamic obstacles like gates and cylinders. The DRL algorithm is based on the Proximal Policy Optimization (PPO) method, which uses an actor-critic neural network architecture to predict drone velocities. Real-time object detection is facilitated by a YOLOv8 model that estimates the positions of obstacles and gates, further refined by an Extended Kalman Filter (EKF) for accurate pose estimation.

Key Results

AgilePilot demonstrates impressive performance in both simulated and real-world tests. The DRL-based approach outperforms classical algorithms like Artificial Potential Field (APF) planners, tripling performance efficiency and achieving more accurate dynamic target tracking. The agent exhibits a success rate of 90% over a series of 75 experiments. In practical experiments, the drones successfully adapt to varying speeds of moving gates and obstacles, achieving velocities up to 3.0 m/s, a significant improvement over baseline methods.

Implications of the Research

AgilePilot's success illustrates the potential of DRL in enhancing autonomous drone capabilities in dynamic environments. The approach offers substantial improvement over traditional methods in terms of adaptability and real-time responsiveness, which are critical for applications in search and rescue operations, delivery services, and surveillance tasks. The ability to tightly integrate perception and decision-making systems in a seamless pipeline exemplifies a significant advancement in UAV autonomy.

Future Directions

The research opens several avenues for future exploration. One potential direction is to enhance the complexity of the simulation environment by including more diverse and unpredictable object dynamics to further stress test the adaptability of the DRL model. Additionally, there is room to explore integrated multi-agent systems, where multiple drones collaborate in real time to accomplish shared objectives. Future research could also focus on reducing the computational overhead of the DRL models, making it feasible to deploy on lighter UAV platforms with limited processing capabilities.

In conclusion, AgilePilot serves as a robust demonstration of how DRL can be harnessed for real-time drone navigation in dynamic settings, setting a foundation for further advancements in the autonomy and efficiency of UAV systems.