Stackelberg Meta-Learning for Strategic Guidance in Multi-Robot Trajectory Planning
share
Guided cooperation is a common task in many multi-agent teaming applications. The planning of the cooperation is difficult when the leader robot has incomplete information about the follower, and there is a need to learn, customize, and adapt the cooperation plan online. To this end, we develop a learning-based Stackelberg game-theoretic framework to address this challenge to achieve optimal trajectory planning for heterogeneous robots. We first formulate the guided trajectory planning problem as a dynamic Stackelberg game and design the cooperation plans using open-loop Stackelberg equilibria. We leverage meta-learning to deal with the unknown follower in the game and propose a Stackelberg meta-learning framework to create online adaptive trajectory guidance plans, where the leader robot learns a meta-best-response model from a prescribed set of followers offline and then fast adapts to a specific online trajectory guidance task using limited learning data. We use simulations in three different scenarios to elaborate on the effectiveness of our framework. Comparison with other learning approaches and no guidance cases show that our framework provides a more time- and data-efficient planning method in trajectory guidance tasks.
READ FULL TEXT