Abstract
Autonomous navigation of micro aerial vehicles (MAVs) in unstructured environments like forest trails presents significant challenges in stability and environmental awareness. This paper presents a robust system utilizing a specialized deep neural network, TrailNet, which predicts both orientation and lateral offset to ensure centered flight. Integrated with real-time object detection and monocular depth estimation on an NVIDIA Jetson TX1, the system achieves high levels of autonomy and smooth control. Experimental results demonstrate 100% autonomy on complex forest trails and stable flights of up to 1 km using off-the-shelf hardware.
Introduction
Navigating a forest trail is a "Goldilocks" problem for robots: they must stay centered enough to avoid thick brush but move fast enough to be useful. Most previous attempts at autonomous trail-following relied solely on predicting the drone's orientation. However, orientation-only systems fail to correct for lateral drift, leading to collisions. Furthermore, many deep learning controllers suffer from "overconfidence," resulting in jerky, oscillatory flight paths. This work addresses these issues by introducing a dual-prediction network and a stability-focused loss function.
System Architecture and Methods
The proposed system runs entirely on an NVIDIA Jetson TX1 mounted on a 3DR Iris+ quadcopter. The software stack is built on ROS and integrates several key components for navigation and awareness.
TrailNet: A Specialized DNN
The core of the navigation system is TrailNet, a modified ResNet-18 architecture optimized for real-time performance and accuracy. Key modifications include the removal of batch normalization and the implementation of Shifted ReLUs (SReLU).
Unlike traditional models that only predict orientation, TrailNet utilizes a 6-category classification scheme. It predicts three categories for orientation (left, center, right) and three for lateral offset (left, center, right relative to the trail midline). This dual approach allows the MAV to actively center itself rather than just pointing in the correct direction.
Stability and Awareness
To prevent jerky flight, a stability-focused loss function incorporating label smoothing and an entropy reward was implemented. This encourages the network to maintain a degree of uncertainty, allowing the controller to blend probabilities for a smoother steering angle. For environmental awareness, the system employs a version of YOLO for detecting people or pets and Direct Sparse Odometry (DSO) for reactive obstacle detection via depth estimation.
Experimental Results
The system was evaluated on challenging forest trails in the Pacific Northwest. In a 250m "zigzag" test, TrailNet achieved 100% autonomy with zero human interventions, outperforming standard ResNet-18 (88%) and previous state-of-the-art models (80%). The MAV successfully completed autonomous flights of up to 1 km in unstructured environments. Disturbance tests confirmed that the dual-prediction model recovers to the trail center significantly faster than orientation-only models. The entire processing stack utilizes approximately 80% of the CPU and 100% of the GPU on the Jetson TX1.
Conclusion and Future Work
This research demonstrates that combining orientation and lateral offset predictions with probabilistic control leads to stable and autonomous MAV navigation in complex environments. While highly effective on defined trails, the system currently struggles in open fields lacking distinct features. Future work will focus on optimizing object detection frame rates and handling complex navigation decisions such as trail switching.
Summary updated