Abstract
How do you learn to navigate an Unmanned Aerial Vehicle (UAV) and avoid obstacles? One approach is to use a small dataset collected by human experts; however, high capacity learning algorithms tend to overfit when trained with little data. An alternative is to use simulation, but the gap between simulation and the real world remains large, especially for perception problems. In this paper, we propose to "bite the bullet" and collect a dataset of crashes itself. We build a drone whose sole purpose is to crash into objects: it samples naive trajectories and crashes into random objects. We crash our drone 11,500 times to create one of the biggest UAV crash datasets. This dataset captures the different ways in which a UAV can crash. We use all this negative flying data in conjunction with positive data sampled from the same trajectories to learn a simple yet powerful policy for UAV navigation. We show that this simple self-supervised model is quite effective in navigating the UAV even in extremely cluttered environments with dynamic obstacles including humans.
Introduction
In the world of robotics, crashing is usually the ultimate failure—the thing engineers spend years trying to avoid. However, the primary reason we lack large-scale real-world data for UAV navigation is this very fear of crashing. Traditionally, researchers have relied on imitation learning from human experts, which is slow and lacks "near-crash" examples, or simulations, which suffer from a significant "reality gap."
By deliberately crashing a drone 11,500 times, researchers from Carnegie Mellon University have developed a self-supervised learning approach that turns failure into a feature. This method allows for the collection of massive amounts of data needed for high-capacity deep learning without the biases of human intervention or the inaccuracies of virtual environments.
The UAV Crash Dataset
The core of this research is a massive, first-of-its-kind UAV Crash Dataset. Using a low-cost Parrot AR-Drone 2.0 equipped with a protective safety hull, the team collected over 40 hours of real-world flight data across 20 different indoor environments, including offices, labs, and corridors.
The data collection process was autonomous: the drone would fly in a straight line until an onboard accelerometer detected an impact. The trajectory was then automatically segmented. Images captured far from the crash were labeled as positive (safe), while images captured just before impact were labeled as negative (unsafe).
Methodology
The researchers employed a standard AlexNet architecture pretrained on ImageNet, modified with a binary softmax classifier to predict crash probability.
At test time, the drone processes three "crops" of its camera view: left, center, and right. If the center crop is predicted to be safe, the drone continues forward. If the center is deemed risky, the model compares the probabilities of the left and right crops to decide the optimal direction to turn. This simple binary classification approach allows the system to scale effectively using well-understood deep learning architectures.
Results and Insights
The "Learning by Crashing" model was evaluated in six challenging environments and compared against depth-prediction baselines and human pilots.
- Superiority over Depth Sensing: Traditional depth-based systems often fail on transparent or untextured surfaces. In a "Glass Door" challenge, the model achieved an average distance of 27.9 meters before crashing, while the depth-based baseline failed almost immediately at 3.1 meters.
- Surpassing Human Performance: In extremely cluttered environments, such as hallways filled with chairs, the drone outperformed human pilots. While humans struggled with control lag and narrow gaps, the model’s training on thousands of failure modes allowed for more robust navigation.
- Visual Cues: The model learns to identify visual cues—like reflections on glass—that signify obstacles, even without an explicit understanding of 3D geometry.
Conclusion
This research demonstrates that "negative data" is a goldmine for building robust robotic systems. By embracing failure, the authors provided a new blueprint for scaling up robot learning in the real world. Future work will look toward expanding the model to handle complex 3D maneuvers and utilizing the "middle-ground" data between safe flight and impact to learn more nuanced behaviors.
Summary updated