Don't Reinvent the Steering Wheel

Laura Yu Zheng, Sanghyun Son, Ming Lin

Primary Area: applications to robotics, autonomy, planning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: trajectory prediction, trajectory forecasting, traffic, driving, kinematics

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We explore using kinematic models as priors for traffic trajectory prediction and show that a simple approach with no additional overhead cost boosts performance in both regular and suboptimal settings.

Abstract:

To make safe and informed decisions, autonomous driving systems can benefit from the capability of predicting the intentions and trajectories of other agents on the road in real-time.
Trajectory forecasting for traffic scenarios has seen great strides in recent years in parallel with advancements in attention-based network architectures and robust, large-scale benchmarks.
However, such models are becoming larger, resource-hungry, and less portable as state-of-the-art pushes for larger-scale of road networks and real-world complexity.
Previous works that achieve state-of-the-art results predict future trajectories as a series of waypoints in Euclidean space, yet do not frame learning through the lenses of classical kinematic models that describe the motion of moving vehicles.
Instead of leaving the network to learn the inherent dynamics of traffic agents, we can instead leverage kinematic models of vehicle dynamics as priors to guide neural networks toward physics-informed solutions earlier in learning.
By combining existing knowledge of how agents move with powerful deep learning techniques, agents learn trajectories that are not only more interpretable but also more plausible in terms of vehicle kinematic constraints.
In this work, we investigate the use of different kinematic formulations as learning priors for trajectory forecasting tasks and evaluate how each affects learning both empirically and analytically.
In addition, we take advantage of time integration in order to derive the original output format of future trajectory coordinates, enabling the use of existing architectures and complementing previous work.
This approach is easy to implement for trajectory forecasting and achieves a considerable performance gain on large-scale benchmarks.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6365