Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

OmniDrones: An Efficient and Flexible Platform for Reinforcement Learning in Drone Control

Botian Xu, Feng Gao, Chao Yu, Ruize Zhang, Yi Wu, Yu Wang

TL;DR

OmniDrones addresses the need for an efficient and flexible RL platform for drone control by building GPU-accelerated, modular simulation on NVIDIA Isaac Sim. The framework offers four drone models, five sensor modalities, four control modes, and more than ten benchmark tasks, along with RL baselines for both single- and multi-agent settings. A bottom-up simulation framework, configurable randomization, and a TorchRL-based learning interface enable scalable experimentation and rapid comparison across algorithms. Empirical results demonstrate high data-throughput (over 10^5 FPS) and illustrate how action-space design and controller choices impact learning performance, paving the way for scalable RL research and sim-to-real transfer in drone systems.

Abstract

In this work, we introduce OmniDrones, an efficient and flexible platform tailored for reinforcement learning in drone control, built on Nvidia's Omniverse Isaac Sim. It employs a bottom-up design approach that allows users to easily design and experiment with various application scenarios on top of GPU-parallelized simulations. It also offers a range of benchmark tasks, presenting challenges ranging from single-drone hovering to over-actuated system tracking. In summary, we propose an open-sourced drone simulation platform, equipped with an extensive suite of tools for drone learning. It includes 4 drone models, 5 sensor modalities, 4 control modes, over 10 benchmark tasks, and a selection of widely used RL baselines. To showcase the capabilities of OmniDrones and to support future research, we also provide preliminary results on these benchmark tasks. We hope this platform will encourage further studies on applying RL to practical drone systems.

OmniDrones: An Efficient and Flexible Platform for Reinforcement Learning in Drone Control

TL;DR

OmniDrones addresses the need for an efficient and flexible RL platform for drone control by building GPU-accelerated, modular simulation on NVIDIA Isaac Sim. The framework offers four drone models, five sensor modalities, four control modes, and more than ten benchmark tasks, along with RL baselines for both single- and multi-agent settings. A bottom-up simulation framework, configurable randomization, and a TorchRL-based learning interface enable scalable experimentation and rapid comparison across algorithms. Empirical results demonstrate high data-throughput (over 10^5 FPS) and illustrate how action-space design and controller choices impact learning performance, paving the way for scalable RL research and sim-to-real transfer in drone systems.

Abstract

In this work, we introduce OmniDrones, an efficient and flexible platform tailored for reinforcement learning in drone control, built on Nvidia's Omniverse Isaac Sim. It employs a bottom-up design approach that allows users to easily design and experiment with various application scenarios on top of GPU-parallelized simulations. It also offers a range of benchmark tasks, presenting challenges ranging from single-drone hovering to over-actuated system tracking. In summary, we propose an open-sourced drone simulation platform, equipped with an extensive suite of tools for drone learning. It includes 4 drone models, 5 sensor modalities, 4 control modes, over 10 benchmark tasks, and a selection of widely used RL baselines. To showcase the capabilities of OmniDrones and to support future research, we also provide preliminary results on these benchmark tasks. We hope this platform will encourage further studies on applying RL to practical drone systems.
Paper Structure (18 sections, 2 equations, 5 figures, 3 tables)

This paper contains 18 sections, 2 equations, 5 figures, 3 tables.

Table of Contents

  1. INTRODUCTION
  2. Related Work
  3. Simulated Environments for drones
  4. Reinforcement learning of drone control
  5. OmniDrones Platform
  6. Simulation Framework
  7. Multi-rotor drone dynamics
  8. Parallelism and scalability
  9. Physical configuration and rigid dynamics
  10. Extending the Drone Models
  11. Randomization
  12. Benchmarking Tasks
  13. Reinforcement Learning with OmniDrones
  14. Experiments
  15. Simulation Performance
  16. ...and 3 more sections

Figures (5)

  • Figure 1: A visualization of the various drone systems in OmniDrones, for which we offer highly efficient simulation, reinforcement learning environments, and benchmarking of baselines.
  • Figure 2: Overview of OmniDrones. OmniDrones provides a foundational library of various sensors and drone models and offers multiple configurations to form diverse drone systems for multifaceted testing. In addition, OmniDrones incorporates several benchmark task logics, enabling the evaluation of the performance of different drone systems across various task objectives. Furthermore, we have implemented and assessed the capabilities of multiple learning algorithms on our benchmark tasks, serving as a baseline for subsequent work.
  • Figure 3: Benchmarking results.
  • Figure 4: Comparison of different drone models on three selected tasks.
  • Figure 5: Comparison of different choices of action space.