Mobile robot navigation in unknown environments requires the ability to quickly interpret sensor data and to respond accordingly. Traditionally, planners have relied on simplifying assumptions, such as planar worlds or point robots. However, adopting such assumptions imposes limitations on the navigation system. Therefore, in order to retain maximum flexibility, different strategies for reducing the computational cost of local planning are needed. Perception space representations improve the efficiency and scaling of local planning. Gaps provide a mechanism to further reduce the complexity of the perceived environment. And machine learning allows a navigation system to learn a policy to dynamically adjust the behavior of a traditional local planning approach in response to the local environment.
The proposed work builds on each of these concepts. Improving the representation capability of perception space representations will enable them to be used in a greater range of scenarios, including for aerial vehicles. Similarly, a gap-based approach for navigating in 3D will enable more efficient navigation for aerial vehicles. Finally, using machine learning to not only tune a planning approach but also to switch between multiple approaches will improve the navigation system’s ability to perform in a wider range of environments.