# UrbanFly: Uncertainty-Aware Planning for Navigation Amongst High-Rises with Monocular Visual-Inertial SLAM Maps

**Authors:** Ayyappa Swamy Thatavarthy, Sudarshan S Harithas, Gurkirat Singh, Arun K Singh, K Madhava Krishna  
**arXiv:** [2204.00865](https://arxiv.org/abs/2204.00865)  
**Date:** 2022-04-02  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2204.00865v1/

## Key points

- UrbanFly is a real-time planning framework for quadrotor navigation in urban high-rise environments
- It plans directly on sparse point clouds from Monocular Visual Inertial SLAM (VINS) using an uncertainty-integrated cuboid representation
- Features two trajectory optimizers that leverage the cuboid representation for enhanced navigation performance
- The first optimizer minimizes collision probability and smoothness cost using a gradient-free cross-entropy method
- The second optimizer uses a sequential convex programming approach based on probabilistic safety estimates
- Both optimizers embed collision violation distributions in Reproducing Kernel Hilbert Space for computational tractability and independence from underlying uncertainty
- UrbanFly outperforms competing baselines in collision rate and trajectory length on a high-fidelity AirSim simulator with synthetic and real-world dataset scenes
- It demonstrates superior navigation performance by exploiting uncertainty formulations that leverage the planarity of the scene, resulting in better safety outcomes and trajectory efficiency compared to models not leveraging structural advantages
- The paper provides an overview of related works, details the pipeline of UrbanFly, and highlights its core components - an uncertainty-integrated environment representation and specialized trajectory optimizers tailored for this representation

## Summary

UrbanFly is a cutting-edge real-time planning framework designed for quadrotor navigation in urban high-rise environments. The framework stands out for its ability to plan directly on sparse point clouds generated by a Monocular Visual Inertial SLAM (VINS) backend. This is made possible thanks to its innovative uncertainty-integrated cuboid representation of the environment. The representation is built using a data-driven monocular plane segmentation network, allowing for faster distance queries compared to traditional voxel-grid representations. UrbanFly features two trajectory optimizers that leverage the uncertainty-integrated cuboid representation in different ways to enhance navigation performance. The first optimizer utilizes a gradient-free cross-entropy method to compute trajectories that minimize collision probability and smoothness cost. Meanwhile, the second optimizer uses a sequential convex programming approach initialized based on probabilistic safety estimates on randomly drawn trajectories. Both trajectory optimizers are designed to be computationally tractable and independent of underlying uncertainty by embedding collision violation distributions in Reproducing Kernel Hilbert Space. This algorithmic innovation enables UrbanFly to outperform competing baselines in metrics such as collision rate and trajectory length on a high-fidelity AirSim simulator augmented with synthetic and real-world dataset scenes. In comparison to existing literature, UrbanFly demonstrates superior navigation performance by exploiting uncertainty formulations that take advantage of the planarity of the scene. This strategic approach results in better safety outcomes and trajectory efficiency compared to models that do not leverage structural advantages when modeling uncertainty over voxel space. The paper is structured as follows: it begins with an overview of related works focusing on monocular SLAM-based planners and chance-constrained optimization techniques. The pipeline of UrbanFly is then detailed, highlighting its core components - an uncertainty-integrated environment representation and specialized trajectory optimizers tailored for this representation. Overall, UrbanFly represents a significant advancement in quadrotor navigation technology, showcasing how innovative approaches to uncertainty-aware planning can lead to improved performance in challenging urban environments.