AirSim Drone Racing Lab (2003.05654v1)

Abstract: Autonomous drone racing is a challenging research problem at the intersection of computer vision, planning, state estimation, and control. We introduce AirSim Drone Racing Lab, a simulation framework for enabling fast prototyping of algorithms for autonomy and enabling machine learning research in this domain, with the goal of reducing the time, money, and risks associated with field robotics. Our framework enables generation of racing tracks in multiple photo-realistic environments, orchestration of drone races, comes with a suite of gate assets, allows for multiple sensor modalities (monocular, depth, neuromorphic events, optical flow), different camera models, and benchmarking of planning, control, computer vision, and learning-based algorithms. We used our framework to host a simulation based drone racing competition at NeurIPS 2019. The competition binaries are available at our github repository.

AirSim Drone Racing Lab: A Framework for Autonomous Drone Racing

The paper "AirSim Drone Racing Lab" presents a sophisticated simulation framework designed to lower the barrier to entry for ML researchers interested in autonomous drone racing. Developed as an extension of Microsoft's AirSim and leveraging the Unreal Engine's graphic capabilities, this framework aims to provide a robust environment for prototyping autonomy algorithms, focusing on perception, planning, state estimation, and control, essential for drone racing.

Overview

Autonomous drone racing is a complex research domain, situated at the intersection of computer vision, planning, state estimation, and control. The AirSim Drone Racing Lab offers a platform that supports the generation of racing tracks in diverse photorealistic environments. It introduces tools for orchestrating drone races, provides a comprehensive suite of gate assets, and accommodates multiple sensor modalities, including monocular, depth, neuromorphic events, and optical flow, alongside various camera models. This framework facilitates benchmarking of planning, control, computer vision, and learning-based algorithms essential for developing autonomous systems.

Simulation Capabilities

The simulation framework is grounded in the hypothesis that simulation can significantly mitigate the complexity associated with real-world autonomous drone experimentation. It is designed with APIs at an abstraction level targeting ML researchers, promoting fast prototyping and verification of the autonomy stack's generalization across varied racing environments. This framework allows researchers to model complex, cluttered environments efficiently—replicating the first person view (FPV) drone racing setup where human pilots navigate with low-resolution, noise-impacted imagery while avoiding obstacles and competing against other drones.

Technical Components

The framework supports the UAV research community by enabling rapid experimentation and testing without financial risks or complexities related to hardware. Notably, AirSim Drone Racing Lab augments the existing capabilities of AirSim by implementing features tailored to drone racing, including dynamic race environments, environment ground truth via voxel grids, API-controlled camera models, and event camera simulation. The architecture facilitates a continuous feedback loop between API control and race state tracking, essential for real-time performance assessments during race conditions.

Race Tracks and Baselines

The paper introduces measures to quantify the complexity of race tracks, such as curvature per unit length for planning and control tasks, and gate visibility for perception tasks. Additionally, baseline algorithms for trajectory planning, tracking, and gate detection are discussed. These baselines serve as benchmarks, enabling researchers to analyze their algorithms against standardized measures of performance in simulated race scenarios.

NeurIPS 2019 Competition

The practical application of this framework was demonstrated during the "Game of Drones" competition at NeurIPS 2019. This competition effectively showcased how simulation can drive engagements and innovations in the ML and robotics communities. The competition consisted of three tiers, each focused on specific aspects of drone racing: planning, perception, and head-to-head competition. Winning teams utilized novel strategies, which included globally optimal trajectory planning, state estimation enhanced by machine learning models, and innovative control policies integrating perception data.

Implications and Future Work

The AirSim Drone Racing Lab represents a significant step forward in making drone racing research accessible to ML communities. By facilitating the development and testing of autonomy algorithms in a risk-free environment, it opens new avenues for innovation in robotics and autonomous systems. The successful integration of realistic simulation environments with complex autonomous behaviors illustrates the potential for advancements in the field. Future developments might focus on enhancing domain randomization for better sim-to-real transfer fidelity, improving event camera simulation, and expanding capabilities to support multi-drone race scenarios.

In conclusion, this paper underscores the utility of realistic simulation frameworks in advancing ML research in autonomous systems, providing researchers with a powerful tool for innovation in drone racing and related areas of robotics.