P4P: Conflict-Aware Motion Prediction for Planning in Autonomous Driving

Qiao Sun, Xin Huang, Brian C. Williams, Hang Zhao

Motion prediction for autonomous vehicles (AVs) has traditionally focused on minimizing geometric errors like Average Displacement Error (ADE). However, this research introduces P4P (Prediction for Planning), a framework that argues for a shift toward "conflict awareness." By prioritizing the identification of potential spatial-temporal conflicts over raw curve-fitting, P4P significantly improves safety in interactive driving scenarios, reducing collision rates from over 12% to just 1.39% in closed-loop simulations.

The Safety-Geometry Gap

Current state-of-the-art (SOTA) motion predictors are often evaluated on how closely their predicted trajectories match recorded ground truth data. While metrics like ADE and FDE provide a measure of geometric accuracy, they fail to capture the downstream impact on planning.

As the authors demonstrate, a prediction can be geometrically accurate but miss a critical interaction, leading the ego vehicle into a collision. Conversely, a "roughly right" prediction that identifies a conflict—even if it is geometrically imprecise—is far more valuable for safe planning. This "Same Error $\neq$ Same Safety" paradox is the core motivation behind P4P.

The paradox of geometric error vs. safety. Two different predictions can result in vastly different safety outcomes for the ego vehicle despite similar error metrics.

The P4P Framework

P4P addresses these shortcomings through a three-stage pipeline designed for robustness and conflict awareness:

  1. Physics-Based Trajectory Generation: Instead of relying solely on deep learning, P4P uses physics-based models to generate consistent, map-aligned candidate trajectories for surrounding agents. This avoids the "hallucinations" or jitter often seen in neural network outputs.
  2. Conflict Search: The system identifies potential intersections in space and time between these candidate trajectories and the ego vehicle's planned path.
  3. Learning-Based Relation Prediction: For every identified conflict, a specialized head (using a VectorNet backbone) classifies the interaction as either Yield or Pass. This high-level semantic understanding allows the planner to make informed decisions about whether to proceed or wait.
The P4P system architecture, illustrating the closed-loop interaction between the simulator, predictor, and planner.

Closed-Loop Evaluation and Results

A key contribution of this work is the use of closed-loop testing. Unlike traditional open-loop benchmarks that simply replay static recorded data, P4P was tested in the InterSim simulator using the Waymo Open Motion Dataset (WOMD). This environment allows traffic agents to react to the ego vehicle's actions in real-time, providing a realistic assessment of interactive safety.

The results highlight a massive gap between traditional metrics and actual safety performance:

  • Conflict Recall: SOTA models like TNT and DenseTNT only identified conflicts correctly about 22-23% of the time. P4P achieved a staggering 97.07% recall.
  • Safety (Collision Rate): In closed-loop simulations, the ego vehicle using SOTA predictors suffered collision rates of 12-14%. Using P4P, the collision rate plummeted to 1.39%.
  • The Trade-off: Interestingly, P4P had much higher ADE/FDE (worse geometric accuracy) than the SOTA models. This confirms that being "roughly right" about a conflict is better than being "precisely wrong" about a trajectory.

Conclusion

P4P demonstrates that simplicity and robustness in trajectory generation, combined with targeted learning for interaction classification, can outperform complex end-to-end models in safety-critical tasks. By focusing on conflicts rather than coordinates, autonomous systems can achieve significantly higher safety margins in the complex, interactive environments of real-world urban driving.

Summary updated

BibTeX
@inproceedings{sun2023p,
  author = {Qiao Sun and Xin Huang and Brian C. Williams and Hang Zhao},
  title = {{P4P: Conflict-Aware Motion Prediction for Planning in Autonomous Driving}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2023},
  doi = {10.1109/iros55552.2023.10342247},
}
Figure