Autonomous navigation in cluttered environments

Abstract

Autonomous navigation in cluttered environments is a necessary capability for emerging robotics applications, from home assistance to disaster rescue and logistics. It remains challenging because of ubiquitous uncertainty, a lack of theoretical guarantees, and high precision control requirements. This thesis studies the navigation problem in cluttered environments for mobile robots using only onboard sensors and without pre-built maps. This process is formulated as an optimization problem with a large number of collision avoidance constraints derived from surrounding obstacles, which is intractable to solve directly. The majority of the research community has converged on the idea of solving this NP-hard problem using model-based or data-driven approaches. Thus, this thesis first presents a data-driven collision avoidance approach for multi-robot systems. By incorporating the concept of reciprocal velocity obstacle (RVO) in the design of observation and reward functions, the training process becomes more stable and encourages the robot to adopt efficient reciprocal local collision avoidance behaviors. Furthermore, considering theoretical guarantees and full-dimensional collision avoidance challenges in previous learning based approaches, we present an accelerated model-based motion planner for cluttered environments. This planner can decompose the complex NP-hard problem into a series of simple convex subproblems and solve them in parallel to generate safe motion commands in real time. Finally, to solve this NP-hard optimization problem in an end-to-end manner, where the raw sensor data is mapped directly to actions, we combine the advantages of data-driven and model-based techniques to construct a model-based learning framework. The model-based module ensures the safety, generalization, and feasibility of the generated trajectory, while the data-driven module can handle hundreds of collision avoidance constraints in real time and improve performance by training with more data. All proposed methods are validated through both simulation and real-world experiments, demonstrating their effectiveness and efficiency.