Deep Kinematic Models for Physically Realistic Prediction of Vehicle Trajectories

Henggang Cui, Thi Nguyen, Fang-Chieh Chou, Tsung-Han Lin, Jeff Schneider, David Bradley, Nemanja DjuricarXiv 2019 · arXiv

Introduction

Predicting the future trajectories of surrounding vehicles is a critical task for autonomous driving systems. While deep learning models have shown great promise in capturing complex data patterns, they often lack an inherent understanding of physical constraints. This "physics-blindness" can result in kinematically infeasible predictions—such as vehicles sliding laterally like a hockey puck or turning with impossible radii—which can lead to unsafe or jerky motion planning for the self-driving vehicle (SDV).

The Deep Kinematic Model (DKM)

To address these issues, researchers at Uber ATG introduced the Deep Kinematic Model (DKM). The core innovation is the integration of a differentiable Kinematic Layer directly into the neural network architecture, forcing the model to obey the laws of physics by design.

Architecture and Methodology

In standard unconstrained models, a backbone network (such as a CNN) predicts a sequence of future $(x, y)$ coordinates directly. In contrast, the DKM backbone is modified to predict control inputs: longitudinal acceleration ($a$) and steering angle ($\gamma$).

These controls are then processed by the Kinematic Layer, which uses a "kinematic bicycle model" to "roll out" the trajectory over time. The layer integrates these controls over a time horizon (e.g., 6 seconds) to produce a physically realistic trajectory consisting of position, heading, and velocity. Because the kinematic equations are fully differentiable, the entire system can be trained end-to-end using standard backpropagation.

Experimental Results

The DKM was evaluated using a massive real-world dataset containing 7.8 million data points collected from 240 hours of driving in various US cities.

  • Physical Feasibility: While unconstrained models produced kinematically impossible trajectories in approximately 26% of cases, the DKM reduced this rate to 0% by design.
  • Accuracy and Heading: DKM achieved the lowest displacement error at a 6-second horizon (4.21m). Notably, it reduced heading (orientation) error by nearly 35% compared to standard models, which is vital for understanding lane-change intentions.
  • Smoothness: Qualitative analysis shows that DKM produces smooth acceleration and turning profiles, avoiding the "noisy" or jittery outputs often seen in coordinate-based predictors.

Limitations and Future Work

While the DKM significantly improves realism, it currently utilizes a kinematic model that assumes no wheel slip. Future iterations could incorporate dynamic models to account for tire forces during extreme maneuvers, such as skidding. Additionally, the model currently assumes a fixed wheelbase; estimating vehicle-specific dimensions (e.g., for buses vs. compact cars) remains an area for future exploration.

Conclusion

By embedding classical kinematic constraints into deep learning architectures, the Deep Kinematic Model bridges the gap between data-driven flexibility and physical reliability. This approach ensures that predicted paths are always physically possible, significantly improving the safety and robustness of autonomous navigation.

Summary updated

BibTeX
@misc{cui2019deep,
  author = {Henggang Cui and Thi Nguyen and Fang-Chieh Chou and Tsung-Han Lin and Jeff Schneider and David Bradley and Nemanja Djuric},
  title = {{Deep Kinematic Models for Physically Realistic Prediction of Vehicle Trajectories}},
  year = {2019},
  eprint = {1908.00219},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1908.00219},
}
Figure