Reproduce “Self-Supervised Learning to Fly using Efficient Semantic Segmentation and Metric Depth Estimation for Low-Cost Autonomous UAVs” on OpenResearch

About this paper

Sebastian Mocanu, Emil Slusanschi, Marius Leordeanu

Mocanu, Slusanschi, and Leordeanu developed a vision-only autonomous flight system for low-cost UAVs, integrating real-time semantic segmentation with a novel adaptive scale factor algorithm for metric depth estimation. Their system achieved a mean distance error of 14.4 cm for metric depth and enabled robust obstacle avoidance with 100% mission success rates for exploration and landing in simulated and real-world tests.

Key insights

• Real-time, accurate metric depth estimation from monocular vision is achievable for low-cost UAVs by leveraging camera parameters and ground plane detection, significantly improving obstacle avoidance.
• Knowledge distillation can effectively transfer complex perception and control policies from robust teacher models to lightweight student networks, enabling deployment on resource-constrained embedded systems.
• Integrating semantic segmentation with metric depth provides a more robust and efficient scene understanding compared to purely semantic-based approaches for autonomous navigation.

Summary by alphaXiv.