Risk-Aware Planning on Point Clouds

Abstract

Navigating complex environments safely is a critical challenge for autonomous drones. This thesis introduces a novel risk-aware planning framework that empowers drones to make smarter, safer decisions. At its core, the framework utilizes an innovative ensemble of neural networks (integrating PointNet, a point cloud processing network, and Gaussian policy-based Multi- Layer Perceptron (MLP) structures) to deliver probabilistic predictions of obstacle distances and their associated uncertainties. Coupled with a jerk-controlled trajectory model, the system leverages Conditional Value-at-Risk (CVaR) and a cross-entropy optimization method to intelligently quantify and mitigate risky trajectories. This allows the drone to confidently navigate by optimally balancing mission objectives against this principled risk measure. Simulation results demonstrate the effectiveness of the proposed framework.