Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TornadoDrone: Bio-inspired DRL-based Drone Landing on 6D Platform with Wind Force Disturbances

Robinroy Peter, Lavanya Ratnabala, Demetros Aschu, Aleksey Fedoseev, Dzmitry Tsetserukou

TL;DR

The paper addresses reliable autonomous drone landings on dynamic platforms under wind disturbances. It proposes TornadoDrone, a bio-inspired TD3-based DRL agent trained in a domain-randomized gym-pybullet environment and validated on Crazyflie 2.1 drones with Vicon indoor localization, outperforming a conventional EKF-PID baseline in both success rate and centimeter-level precision. The results demonstrate the model's ability to infer and compensate for wind effects from flight dynamics rather than direct wind measurements, achieving high robustness across static, linear, and complex platform motions. This work advances autonomous drone operations in dynamic, wind-prone settings with potential impact on surveillance, search-and-rescue, and emergency response tasks.

Abstract

Autonomous drone navigation faces a critical challenge in achieving accurate landings on dynamic platforms, especially under unpredictable conditions such as wind turbulence. Our research introduces TornadoDrone, a novel Deep Reinforcement Learning (DRL) model that adopts bio-inspired mechanisms to adapt to wind forces, mirroring the natural adaptability seen in birds. This model, unlike traditional approaches, derives its adaptability from indirect cues such as changes in position and velocity, rather than direct wind force measurements. TornadoDrone was rigorously trained in the gym-pybullet-drone simulator, which closely replicates the complexities of wind dynamics in the real world. Through extensive testing with Crazyflie 2.1 drones in both simulated and real windy conditions, TornadoDrone demonstrated a high performance in maintaining high-precision landing accuracy on moving platforms, surpassing conventional control methods such as PID controllers with Extended Kalman Filters. The study not only highlights the potential of DRL to tackle complex aerodynamic challenges but also paves the way for advanced autonomous systems that can adapt to environmental changes in real-time. The success of TornadoDrone signifies a leap forward in drone technology, particularly for critical applications such as surveillance and emergency response, where reliability and precision are paramount.

TornadoDrone: Bio-inspired DRL-based Drone Landing on 6D Platform with Wind Force Disturbances

TL;DR

The paper addresses reliable autonomous drone landings on dynamic platforms under wind disturbances. It proposes TornadoDrone, a bio-inspired TD3-based DRL agent trained in a domain-randomized gym-pybullet environment and validated on Crazyflie 2.1 drones with Vicon indoor localization, outperforming a conventional EKF-PID baseline in both success rate and centimeter-level precision. The results demonstrate the model's ability to infer and compensate for wind effects from flight dynamics rather than direct wind measurements, achieving high robustness across static, linear, and complex platform motions. This work advances autonomous drone operations in dynamic, wind-prone settings with potential impact on surveillance, search-and-rescue, and emergency response tasks.

Abstract

Autonomous drone navigation faces a critical challenge in achieving accurate landings on dynamic platforms, especially under unpredictable conditions such as wind turbulence. Our research introduces TornadoDrone, a novel Deep Reinforcement Learning (DRL) model that adopts bio-inspired mechanisms to adapt to wind forces, mirroring the natural adaptability seen in birds. This model, unlike traditional approaches, derives its adaptability from indirect cues such as changes in position and velocity, rather than direct wind force measurements. TornadoDrone was rigorously trained in the gym-pybullet-drone simulator, which closely replicates the complexities of wind dynamics in the real world. Through extensive testing with Crazyflie 2.1 drones in both simulated and real windy conditions, TornadoDrone demonstrated a high performance in maintaining high-precision landing accuracy on moving platforms, surpassing conventional control methods such as PID controllers with Extended Kalman Filters. The study not only highlights the potential of DRL to tackle complex aerodynamic challenges but also paves the way for advanced autonomous systems that can adapt to environmental changes in real-time. The success of TornadoDrone signifies a leap forward in drone technology, particularly for critical applications such as surveillance and emergency response, where reliability and precision are paramount.
Paper Structure (21 sections, 9 equations, 9 figures, 5 tables)

This paper contains 21 sections, 9 equations, 9 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Methodology
  4. Simulation Environment Setup
  5. Deep Reinforcement Learning Framework
  6. TornadoDrone Agent Architecture
  7. Reward Function
  8. Training Protocol
  9. Real-World Validation Setup
  10. Indoor Localization System (Vicon)
  11. Crazyflie Drones
  12. Baseline Comparison
  13. Crazyflie On-board PID Controller with EKF
  14. Experiments
  15. Experimental Design
  16. ...and 6 more sections

Figures (9)

  • Figure 1: Composite frame illustrating key phases of autonomous landing: (a) Complete trajectory overview, (b) Instant re-planning in response to external forces, (c) Sudden recovery behavior (d) Adaptation to sudden directional changes of the moving landing platform.
  • Figure 2: Simulation setup in gym-pybullet environment.
  • Figure 3: Architecture of the DRL model illustrating inputs, hidden layers, and action mechanisms.
  • Figure 4: Origin view of TornadoDrone's reward function emphasizing safety and behavior.
  • Figure 5: Mean reward vs training steps, showcasing learning progress.
  • ...and 4 more figures