The ability to navigate in dynamic environments has become an urgent demand for robots working around humans. Challenges persist in robot navigation in dense crowds or other dynamic environments. Whether it is achieved through a modular or end-to-end pipeline, navigation in such environments suffers from the uncertainty of dynamic objects and the variability of complex contexts. This thesis focuses on these factors to realize efficient and effective navigation systems.

Navigation systems in low dynamic environments are first investigated. Conventional navigation systems that follow the sense-plan-control pipeline tackle dynamic obstacles with time-inefficient replanning or computationally demanding spatial-temporal optimization. This thesis proposes a hierarchical trajectory planning me...[ Read more ]

The ability to navigate in dynamic environments has become an urgent demand for robots working around humans. Challenges persist in robot navigation in dense crowds or other dynamic environments. Whether it is achieved through a modular or end-to-end pipeline, navigation in such environments suffers from the uncertainty of dynamic objects and the variability of complex contexts. This thesis focuses on these factors to realize efficient and effective navigation systems.

Navigation systems in low dynamic environments are first investigated. Conventional navigation systems that follow the sense-plan-control pipeline tackle dynamic obstacles with time-inefficient replanning or computationally demanding spatial-temporal optimization. This thesis proposes a hierarchical trajectory planning method that decouples the spatial planning and temporal planning to reduce computation time and memory usage. End-to-end visual navigation suffers from performance degradation caused by domain shift or dataset bias. This thesis proposes to use human attention to enhance the generalization capability of the end-to-end autonomous driving network. By selectively filtering out task-irrelevant information, like the changeful background, the method shows significantly better performance and lower model uncertainty in unseen environments than the baseline.

In high dynamic environments, interaction modeling becomes essential for navigation tasks. Graph representation is utilized to model the interactions among robots and humans effectively. In collision avoidance, the crowd feature is aggregated through the graph convolutional network (GCN), where the robot assigns different attention to the neighborhood through the adjacency matrix. In pedestrian trajectory prediction, coherent motion in the crowd is fully leveraged for interaction modeling. The intergroup and intragroup interactions are captured separately through two graphs to give accurate and realistic predictions. The experimental results prove the superiority of the methods over the state-of-the-art methods in terms of navigation success rate and trajectory prediction error. These modules can be further deployed to facilitate navigation in high dynamic environments.