A Framework for Offline Risk-aware Planning of Low-altitude Aerial Flights during Urban Disaster Response

Abstract

Disaster response missions are dynamic and dangerous events for first responders. Active situational awareness is critical for effective decision-making, and unmanned aerial assets have successfully extended the range and output of sensors. Aerial assets have demonstrated their capability in disaster response missions via decentralized operations. However, literature and industry lack a systematic investigation of the algorithms, datasets, and tools for aerial system trajectory planning in urban disasters that optimizes mission performance and guarantee mission success. This work seeks to develop a framework and software environment to investigate the requirements for offline planning algorithms and flight risk models when applied to aerial assets exploring urban disaster zones. This is addressed through the creation of rapid urban maps, efficient flight planning algorithms, and formal risk metrics that are demonstrated in scenario-driven experiments using Monte Carlo simulation. First, rapid urban mapping strategies are independently compared for efficient processing and storage through obstacle and terrain layers. Open-source data is used when available and is supplemented with an urban feature prediction model trained on satellite imagery using deep learning. Second, sampling-based planners are evaluated for efficient and effective trajectory planning of nonlinear aerial dynamic systems. The algorithm can find collision-free, kinodynamic feasible trajectories using random open-loop control targets. Alternative open-loop control commands are formed to improve the planning algorithm’s speed and convergence. Third, a risk-aware implementation of the planning algorithm is developed that considers the uncertainty of energy, collisions, and onboard viewpoint data and maps them to a single measure of the likelihood of mission failure. The three modules are combined in a framework where the rapid urban maps and risk-aware planner are evaluated against benchmarks for mission success, performance, and speed while creating a unique set of benchmarks from open-source data and software. One, the rapid urban map module generates a 3D structure and terrain map within 20 meters of data and in less than 5 minutes. The Gaussian Process terrain model performs better than B-spline and NURBS models in small-scale, mountainous environments at 10-meter squared resolution. Supplementary data for structures and other urban landcover features is predicted using the Pix2Pix Generative Adversarial Network with a 3-channel encoding for nine labels. Structures, greenspaces, water, and roads are predicted with high accuracy according to the F1, OIU, and pixel accuracy metrics. Two, the sampling-based planning algorithm is selected for forming collision-free, 3D offline flight paths with a black-box dynamics model of a quadcopter. Sampling-based planners prove successful for efficient and optimal flight paths through randomly generated and rapid urban maps, even under wind and noise uncertainty. The Stable-Sparse-RRT, SST, algorithm is shown to improve trajectories for minimum Euclidean distance more consistently and efficiently than the RRT algorithm, with a 50% improvement in finite-time path convergence for large-scale urban maps. The forward propagation dynamics of the black-box model are replaced with 5-15 times more computationally efficient motion primitives that are generated using an inverse lower-order dynamics model and the Differential Dynamic Programming, DDP, algorithm. Third, the risk-aware planning algorithm is developed that generates optimal paths based on three risk metrics of energy, collision, and viewpoint risk and quantifies the likelihood of worst-case events using the Conditional-Value-at-Risk, CVaR, metric. The sampling-based planning algorithm is improved with informative paths, and three versions of the algorithm are compared for the best performance in different scenarios. Energy risk in the planning algorithm results in 5-35% energy reduction and 20-30% more consistency in finite-time convergence for flight paths in large-scale urban maps. All three risk metrics in the planning algorithm generally result in more energy use than the planner with only energy risk, but reduce the mean flight path risk by 10-50% depending on the environment, energy available, and viewpoint landmarks. A final experiment in an Atlanta flooding scenario demonstrates the framework’s full capability with the rapid urban map displaying essential features and the trajectory planner reporting flight time, energy consumption, and total risk. Furthermore, the simulation environment provides insight into offline planning limitations through Monte Carlo simulations with environment wind and system dynamics noise. The framework and software environment are made available to use as benchmarks in the field to serve as a foundation for increasing the effectiveness of first responders’ safety in the challenging task of urban disaster response.