Combining Auxiliary Losses for Safer and More Robust Trajectory Prediction

Ahmad Rahimi, Alexandre Alahi

Keywords: Vehicle Trajectory Prediction, Multi-loss training, Loss Balancing, Autonomous Vehicles

TL;DR: We introduce auxiliary losses and an adaptive weighting scheme that improve road adherence and diversity in vehicle trajectory prediction.

Abstract:

Accurate trajectory prediction is essential for the safety and reliability of autonomous systems. Despite recent progress, models still struggle with scene compliance, often producing off-road or traffic-violating forecasts. We revisit and enhance three intuitive auxiliary objectives—Offroad Loss, Direction Consistency Loss, and Diversity Loss—that enhance map adherence, traffic rule compliance, and trajectory coverage. While each improves a specific aspect, our key finding is that only their combination delivers robust road-compliant predictions. To make this practical, we propose a lightweight adaptive weighting scheme that balances auxiliary losses automatically, succeeding where existing multi-task training strategies fail.
Extensive experiments on nuScenes and Argoverse 2 show consistent gains in safety and robustness without sacrificing accuracy, with 43\% decrease in off-road errors on average. Notably, under the SceneAttack benchmark, which perturbs road geometry to create out-of-distribution driving scenarios, our method reduces off-road errors by 25\%, demonstrating that learned road compliance transfers to unseen environments. Our plug-and-play package can be integrated into any trajectory predictor, and code will be released.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 12478