Learning to Play Trajectory Games Against Opponents With Unknown Objectives

Many autonomous agents, such as intelligent vehicles, are inherently required to interact with one another. Game theory provides a natural mathematical tool for robot motion planning in interactive settings. However, tractable algorithms problems usually rely on strong assumption, namely that the objectives of all players scene known. To make tools applicable ego-centric only local information, we propose an adaptive model-predictive game solver, which jointly infers other players' online and computes corresponding generalized Nash equilibrium (GNE) strategy. The adaptivity our approach is enabled by differentiable trajectory solver whose gradient signal used maximum likelihood estimation (MLE) opponents' objectives. This differentiability pipeline facilitates direct integration elements, neural networks (NNs). Furthermore, contrast existing solvers cost inference games, method handles not partial state observations but also general inequality constraints. In two simulated traffic scenarios, find superior performance over both game-theoretic methods non-game-theoretic control (MPC) approaches. We demonstrate approach's real-time capabilities robustness two-player hardware experiments.