Introduction
In the development of autonomous driving systems, two tasks are paramount: trajectory prediction—anticipating the movement of surrounding vehicles—and trajectory generation—creating realistic traffic scenarios for simulation and testing. Traditionally, these tasks are treated as distinct problems. Existing prediction models often function as "black boxes," providing displacement coordinates without explaining the underlying driver intent. Conversely, generative models frequently struggle to produce specific, safety-critical scenarios because they lack fine-grained control over agent behaviors.
This paper introduces the Trajectory Autoencoder (TAE), a unified framework designed to bridge this gap. By explicitly modeling high-level behaviors such as "aggressiveness" and "intention" within a semi-supervised adversarial autoencoder, TAE allows researchers to manipulate driver behavior to generate diverse and controllable scenarios for testing autonomous planners.
Methodology
Context Modeling
To accurately predict or generate a path, the model must first understand the environmental context. TAE employs a multi-modal feature extraction pipeline:
- Temporal Features: 1-D Convolutional Neural Networks (CNNs) process the historical trajectories of vehicles.
- Spatial and Social Features: Graph Convolutional Networks (GCNs) encode the road map geometry and vehicle-to-vehicle interactions.
- Fusion: An attention mechanism fuses these disparate features into a unified 128-dimensional context vector.
The Structured Latent Space
Unlike standard autoencoders that utilize an uninterpretable latent space, TAE partitions its latent representation into three physically meaningful components:
- Intention ($z_{intent}$): A categorical distribution representing high-level maneuvers: moving forward, turning/changing lanes left, or turning/changing lanes right.
- Aggressiveness ($z_{agg}$): A one-dimensional value following a log-normal distribution. The authors use "time headway" (the temporal gap between vehicles) as a proxy for driving style.
- Gaussian Residual ($z_{gauss}$): A 6-dimensional space that captures stochastic variations and nuances not covered by intention or style.
Semi-supervised Adversarial Training
Recognizing that labels for "aggressiveness" are rarely available in large-scale datasets, the authors utilize a semi-supervised approach. The model is trained using a small subset of labeled data to ground the latent variables, while an adversarial loss ensures that the generated latent vectors adhere to the specified distributions (Gaussian, Log-normal, and Categorical).
Key Insights
Modeling Driver Style
Through analysis of the Argoverse motion forecasting dataset, the authors discovered that time headway—a critical metric for aggressiveness—aligns more closely with a log-normal distribution than a standard Gaussian one. Integrating this domain knowledge directly into the architecture allows for more realistic behavioral modeling.
Disentanglement and Controllability
By separating intention from aggressiveness, the model achieves high levels of disentanglement. This allows a user to generate a "conservative lane change" versus an "aggressive lane change" simply by adjusting the values in the structured latent space, providing a powerful tool for stress-testing planners.
Results and Discussion
The TAE model was evaluated using the Argoverse dataset, featuring over 30,000 real-world scenarios from Miami and Pittsburgh.
- Diversity in Generation: In diversity benchmarks, TAE produced 5x more diverse scenarios than a vanilla Adversarial Autoencoder (AAE) at high thresholds, significantly outperforming GCN and Transformer-based baselines.
- Behavioral Accuracy: The model achieved an 89.16% accuracy in predicting driver intention and an MSE of 0.36 for aggressiveness estimation.
- Competitive Prediction: Despite its focus on controllability, TAE remains a high-performing predictor. It achieved an Average Displacement Error (ADE) of 1.73m, placing it in direct competition with state-of-the-art models like LaneGCN (1.71m) and TNT (1.77m).
- Safety-Critical Scenario Generation: By manually shifting the "aggressiveness" variable, the authors demonstrated the ability to generate 227.5% more risky scenarios (including collisions and near-misses) compared to baseline models.
Conclusion
TAE represents a significant step toward interpretable and controllable motion modeling. While there is a minor trade-off in raw displacement accuracy due to the information loss inherent in low-dimensional latent mapping, the gain in controllability is substantial. Future work will focus on modeling more complex interactions, such as yielding and pedestrian dynamics, to further enhance the realism of generated traffic environments.
Summary updated