Learning Monocular Reactive UAV Control in Cluttered Natural Environments (1211.1690v1)

Abstract: Autonomous navigation for large Unmanned Aerial Vehicles (UAVs) is fairly straight-forward, as expensive sensors and monitoring devices can be employed. In contrast, obstacle avoidance remains a challenging task for Micro Aerial Vehicles (MAVs) which operate at low altitude in cluttered environments. Unlike large vehicles, MAVs can only carry very light sensors, such as cameras, making autonomous navigation through obstacles much more challenging. In this paper, we describe a system that navigates a small quadrotor helicopter autonomously at low altitude through natural forest environments. Using only a single cheap camera to perceive the environment, we are able to maintain a constant velocity of up to 1.5m/s. Given a small set of human pilot demonstrations, we use recent state-of-the-art imitation learning techniques to train a controller that can avoid trees by adapting the MAVs heading. We demonstrate the performance of our system in a more controlled environment indoors, and in real natural forest environments outdoors.

Monocular Reactive UAV Control in Cluttered Environments

The paper "Learning Monocular Reactive UAV Control in Cluttered Natural Environments" explores the challenges and advancements in autonomous navigation for Micro Aerial Vehicles (MAVs), particularly in environments densely populated with obstacles, such as forests. While large Unmanned Aerial Vehicles (UAVs) benefit from carrying heavy sensors like radars and lidars, MAVs face significant limitations due to their reduced payload capacities. The authors address this gap by proposing a monocular vision-based navigation system utilizing a passive sensor—a single camera. Their contributions can be encapsulated as follows:

Methodology Overview

The research leverages imitation learning, specifically the DAgger (DAtaset Aggregation) algorithm, to train MAV controllers. Unlike traditional supervised learning, this iterative technique iterates over data distribution changes induced by the learner's actions. Initially, a human expert pilots the MAV to demonstrate optimal navigation paths. During subsequent iterations, the system refines its policies by collecting data in scenarios it would encounter autonomously, paralleling the expert's decisions for corrective feedback.

The controller contextualizes its decision-making process within a rich feature space extracted from camera images. These features include Radon transform statistics, structure tensor statistics, Laws’ masks, and optical flow, providing a robust set of visual cues. The training comprises various canonical obstacle arrangements in a controlled environment before deploying the MAV in dense forest settings with increased complexity.

Experimental Validation

Validation of the proposed method occurs in both indoor and outdoor settings. Indoor tests leverage a motion capture arena with artificial obstacles, aiding quantitative analysis. In this environment, DAgger iteratively improves its performance, achieving success in obstacle avoidance after a few iterations. Qualitative assessments in real-world settings, conducted in low and high-density forest areas, demonstrate the system's capacity to maintain sustained navigation with minimal failures attributed primarily to the field of view limitations rather than the decision-making process.

Implications and Future Work

This paper presents key implications both practically and academically. The paper enhances MAV autonomy, potentially expanding their utility in cluttered environments where traditional sensors are impractical. The exploration of imitation learning techniques like DAgger in MAV control scenarios positions this work as a stepping stone toward integrating more advanced planning capabilities.

Future research directions may include addressing the observable constraints such as the limited field of view and latency issues. Additionally, incorporating higher-level planning systems, possibly through receding horizon methods or enhanced mapping capabilities, could complement the reactive control layer.

Ultimately, this research underscores the potential of cost-effective, visual-based navigation systems for MAVs, opening avenues for deployment in various applications ranging from environmental monitoring to search and rescue missions in complex environments. The introduction of memory-aware imitation learning strategies might further enhance this system's robustness and efficiency, paving the way for more sophisticated autonomous aerial platforms.