GoIRL: Graph-Oriented Inverse Reinforcement Learning for Multimodal Trajectory Prediction

Muleilan Pei, Shaoshuai Shi, Lu Zhang, Peiliang Li, Shaojie Shen

TL;DR: A novel trajectory predictor integrating the inverse reinforcement learning paradigm with vectorized context representations.

Abstract: Trajectory prediction for surrounding agents is a challenging task in autonomous driving due to its inherent uncertainty and underlying multimodality. Unlike prevailing data-driven methods that primarily rely on supervised learning, in this paper, we introduce a novel **G**raph-**o**riented **I**nverse **R**einforcement **L**earning (GoIRL) framework, which is an IRL-based predictor equipped with vectorized context representations. We develop a feature adaptor to effectively aggregate lane-graph features into grid space, enabling seamless integration with the maximum entropy IRL paradigm to infer the reward distribution and obtain the policy that can be sampled to induce multiple plausible plans. Furthermore, conditioned on the sampled plans, we implement a hierarchical parameterized trajectory generator with a refinement module to enhance prediction accuracy and a probability fusion strategy to boost prediction confidence. Extensive experimental results showcase our approach not only achieves state-of-the-art performance on the large-scale Argoverse & nuScenes motion forecasting benchmarks but also exhibits superior generalization abilities compared to existing supervised models.

Lay Summary: Accurately predicting how surrounding traffic participants, such as cars, cyclists, and pedestrians, will move is essential for safe autonomous driving, yet remains a major challenge due to the inherently multi-modal nature of human behavior. Most existing approaches rely on supervised learning, which tends to overfit to seen behaviors and struggles to generalize to new or unseen driving scenarios. To address this, we propose GoIRL, a novel trajectory prediction framework built on Inverse Reinforcement Learning (IRL), a technique that infers the goals and intentions behind observed movements, rather than merely imitating them. GoIRL leverages a vectorized lane-graph representation to capture rich road context and introduces a feature adaptor that bridges this information into a grid-based format compatible with the IRL paradigm. This enables the model to learn reward functions that guide realistic and diverse future motion plans. We further develop a hierarchical trajectory generator with refinement modules that boost both prediction accuracy and confidence. GoIRL achieves state-of-the-art performance on large-scale motion forecasting benchmarks and demonstrates superior generalization to changes in drivable areas than current supervised models. Our research represents a significant advancement towards enhanced and generalizable foresight, which is critical for the safe deployment of autonomous vehicles in the real world.

Primary Area: Applications->Computer Vision

Keywords: Autonomous Driving, Trajectory Prediction, Inverse Reinforcement Learning

Submission Number: 11512