ExpliDrive: Bridging Model Predictive Control and Transformers for Interactive Autonomous Driving

Abstract

Autonomous driving (AD) continues to grapple with the complexity of dynamic and interactive traffic environments, where the primary difficulty stems from insufficient modeling of inter-vehicle interactions—particularly, how autonomous agents should perceive and respond to surrounding vehicles’ influence. To address this, this paper proposed ExpliDrive, an explainable data-driven approach for interaction-aware autonomous driving. Its highlights lie in bridging Model Predictive Control (MPC) and Transformers. The proposed approach builds a generalized system dynamics in which interaction effects between vehicles are explicitly modeled. Specifically, a Transformer encoder-decoder is employed to encode the interaction patterns among vehicles, and these learned effects are seamlessly embedded into the motion planning process. Hence, the proposed approach bears following features: i) enabling proactively interaction-aware autonomous driving; ii) data-driven yet explainable; iii) integrating the prediction into motion planning. Open-looped evaluation demonstrates the proposed approach achieves the lowest prediction errors, from ADE@1s (0.16m) to ADE@5s (0.80m). Close-looped planning shows the proposed approach has significant benefits in driving success rate and flexibility.