Conditional Prediction by Simulation for Automated Driving

Fabian Konstantinidis^1,2, Moritz Sackmann¹, Ulrich Hofmann¹, Christoph Stiller²

¹CARIAD SE, ²Karlsruhe Institute of Technology
FAS Workshop 2025

Prediction setup: We propose an autoregressive prediction model that enables conditional predictions by learning a reactive behavior policy for simulating surrounding vehicles in a closed-loop simulation. Due to the stepwise rollout of the prediction, vehicles can respond to each other's movements in the subsequent simulation step.

Abstract

Modular automated driving systems commonly handle prediction and planning as sequential, separate tasks, thereby prohibiting cooperative maneuvers. To enable cooperative planning, this work introduces a prediction model that models the conditional dependencies between trajectories. For this, predictions are generated by a microscopic traffic simulation, with the individual traffic participants being controlled by a realistic behavior model trained via Adversarial Inverse Reinforcement Learning. By assuming various candidate trajectories for the automated vehicle, we generate predictions conditioned on each of them. Furthermore, our approach allows the candidate trajectories to adapt dynamically during the prediction rollout. We demonstrate the effectiveness of our method in various traffic scenarios with distinct road layouts and traffic densities.

Training Setup

We employ Adversarial Inverse Reinforcement Learning (AIRL) [1] to learn the behavior model. AIRL combines Reinforcement Learning (RL) with the ideas of Generative Adversarial Networks [2] and applies them to the task of Imitation Learning. Specifically, the behavior model is trained via RL maximizing a reward signal, with the reward signal being approximated by a discriminator model. The discriminator is trained to assign higher scores to more realistic samples.

Quantitative Prediction Results

Normalized histogram of executed actions of the learned behavior model (blue) and the corresponding ground truth vehicles (orange). The large overlap indicates that the policy effectively captures the correct distribution of actions. The peak at 0 m/s² occurs due to the preprocessing, where for vehicles at standstill (e.g., waiting at a yield line) both their acceleration and steering angle were set to zero.

Prediction performance after 10s: Each cell shows the mean and standard deviation of seven models trained with different random seeds. The BC models were trained by minimizing the negative log-likelihood of the real-world drivers' actions under the predicted action distribution. The results show that via AIRL realistic beahvior models are obtained that, when executed in a closed-loop simulation, generate accurate and scene-consistent predictions.

Conditional Prediction: Example Situations

To demonstrate that the proposed prediction model is capable of making conditional predictions, we modify the planned trajectories of individual vehicles and predict the remaining vehicles. Similarly, an automated vehicle could query the prediction model with different planned trajectories for itself and select a plan based on the corresponding predictions.

Example 1: Unsignalized Intersection with All-Way Stop

In this scenario, the pink vehicle (ID 6) intends to turn right, while the oncoming red vehicle (ID 3) intends to turn left towards the same exit. Assuming that ID 6 is an automated vehicle, it must decide whether to proceed before or yield to ID 3. We predict the evolution of the traffic situation for both options separately to evaluate the consequences.

Ground truth evolution.

Conditional prediction: "What if #6 (pink) drives before #3 (red)?".

Conditional prediction: "What if #6 (pink) gives way to #3 (red)?".

Example 2: Roundabout

In this scenario, the green vehicle (ID 2) should yield to the blue vehicle (ID 0). To showcase the prediction model’s reactiveness, we also predict the traffic evolution if ID 2 enters the roundabout, cutting off ID 0.

Ground truth evolution.

Conditional prediction: "What if #2 (green) gives way #0 (blue)?".

Conditional prediction: "What if #2 (green) cuts #0 (blue) off?".

Example 3: Merging

In this scenario, the gray vehicle (ID 7) is trying to merge onto the left lane, which requires cooperation from the yellow vehicle (ID 8). The predictions below show ID 8 either cooperating in one case or not in the other.

Ground truth evolution.

Conditional prediction: "What if #8 (yellow) is cooperative and allows #7 (gray) to merge?".

Conditional prediction: "What if #8 (yellow) is not cooperative and accelerates to prevent #7 (gray) from merging?".

Acknowledgment

The authors would like to thank Matthias Dingwerth and Matthias Steiner for their support and contributions to the development of the simulation framework.

BibTeX


              @article{konstantinidis2025conditional,
                title={Conditional Prediction by Simulation for Automated Driving},
                author={Konstantinidis, Fabian and Sackmann, Moritz and Hofmann, Ulrich and Stiller, Christoph},
                journal={arXiv preprint arXiv:2502.03286},
                year={2025}
              }