Robots that autonomously navigate real-world 3D cluttered environments need to safely traverse terrain with abrupt changes in surface normal and elevation. This thesis presents a decentralized robot architecture for navigation and a novel sim-to-real pipeline for learning real-world navigation in simulation using deep reinforcement learning.

The decentralized robot architecture design avoids the requirement of a powerful central server and improves the robustness of the processes against hardware and software failures by isolating the computing units. A set of experiments were conducted to demonstrate the robustness of the architecture's mapping system against challenging environmental conditions.

The sim-to-real pipeline uses deep reinforcement learning to learn a navigation policy from data collected in simulation. It incorporated a combination of sim-to-real strategies to address the reality gap that uniquely exists for 3D navigation problems. A set of real-world experiments demonstrated that the pipeline successfully transferred the learned navigation policy into the real world.