Analyzing and learning diverse behaviors is pivotal in advancing embodied AI, particularly in the realms of robotics and autonomous driving. This dissertation explores three critical aspects of behavior-related research: prediction, generation, and skill learning.
The research begins by addressing the interactive behavior prediction problem in driving scenarios. It employs probabilistic graphical methods to interpret and model the intention changes of vulnerable road users, providing trajectory predictions in interactive scenarios. It then introduces a learning-based approach that leverages domain-specific knowledge to facilitate joint prediction for vehicle interactions, offering interpretable predictions of multi-modal interactive trajectories.
Subsequently, the focus shifts to modeling and generating interactive behaviors. This includes introducing a generative model for learning conditional trajectory generation in joint interactions from collected datasets, with capabilities for generating critical interactions through controllable parameters in provided road scenarios. Further, the work extends to more generalized and complex indoor scenarios where agents are controlled in distributed settings without communication. Potential games are used to model collaborative behaviors between humans and robots, and online optimizations are used to simulate human-like interactions in challenging scenarios. This framework not only generates diverse interactions but also serves to evaluate navigation algorithms.
The final part of the dissertation explores different methods for learning behavioral skills. This includes a parameter compositional framework that utilizes multi-task reinforcement learning and transfer learning to acquire generalized manipulation skills efficiently. An adaptive energy reward design is then detailed, aiding in natural locomotion behavior learning across various speeds and gaits in quadrupedal robots. Moreover, a generalized framework employing large language models addresses partially observable tasks in robotics, showcasing the utility of reinforcement and supervised learning across diverse behavioral contexts.
Overall, this dissertation integrates an array of innovative approaches for predicting, generating, and learning behaviors within autonomous systems, advancing the field of embodied AI. These contributions extend the theoretical understanding of complex behavioral dynamics and enhance practical implementations in real-world applications. By introducing robust, scalable, and interpretable models and algorithms, this dissertation aims to increase the adaptability and efficiency of robotic systems across diverse operational environments.
