Abstract
PITA (Physics-Informed Trajectory Autoencoder) is a novel architecture designed to generate physically realistic vehicle trajectories by embedding kinematic constraints directly into the neural network's loss function. Traditional generative models often produce "jittery" paths that, while spatially accurate, require impossible physical maneuvers. By incorporating a Kinematic Bicycle Model (KBM) and a hybrid loss function, PITA produces smooth trajectories that align with real-world vehicle dynamics. Evaluated on the rounD dataset, PITA demonstrates a superior balance between reconstruction accuracy and physical plausibility, providing a high-fidelity foundation for simulating edge cases in autonomous driving.
Introduction
The safe deployment of autonomous vehicles (AVs) necessitates rigorous testing against "edge cases"—rare, high-risk scenarios that are underrepresented in real-world driving data. To address this, researchers often employ generative models like autoencoders to augment datasets by synthesizing new trajectories from a learned latent space.
However, standard autoencoders suffer from a persistent "jitter" problem. Typically trained to minimize Mean Squared Error (MSE) between input and output coordinates, these models treat trajectory points as independent entities. The resulting paths often exhibit high-frequency noise that implies physically impossible behavior, such as instantaneous, infinite acceleration. Such low-fidelity data is unsuitable for high-stakes simulation, as it fails to represent the actual capabilities of a vehicle.
Methodology: Physics-Informed Learning
The authors propose PITA, an architecture that addresses the jitter problem by integrating a Physics-Informed Loss Function into a standard 12-layer autoencoder.
Kinematic Constraints
PITA utilizes a 3-DOF Kinematic Bicycle Model (KBM) to track position ($x, y$), heading ($\theta$), and velocity ($v$). The model treats curvature ($\kappa$) and acceleration ($a$) as control inputs. During training, the total loss is calculated as a weighted sum of a standard reconstruction loss and a physical loss that penalizes deviations from the KBM differential equations.
Training Dynamics
To prevent model collapse—where the network might predict a stationary state to trivially satisfy physical laws—the researchers implemented an exponential Loss Scheduling scheme. The physics penalty starts at a low value and increases as training progresses, allowing the model to first learn the general shape of the trajectory before refining its physical consistency.
The study compares PITA against two baselines: a Simple Autoencoder (pure MSE) and an Action-Space Autoencoder (which predicts control inputs and integrates them, ensuring physical validity at the cost of reconstruction accuracy).
Experimental Setup
PITA was evaluated using the rounD dataset, which features drone-recorded trajectories of vehicles in German roundabouts. Roundabouts offer a challenging testbed due to their complex, non-linear motion patterns. The dataset consists of 13,509 trajectories, each spanning a 14-second window (350 time steps) sampled at 25 Hz.
Results and Discussion
The evaluation focused on the trade-off between reconstruction accuracy and physical realism.
Smoothness and Realism
Qualitatively, PITA variants produce significantly smoother paths than the Simple Autoencoder. Quantitatively, the Simple Autoencoder required unrealistically high control inputs to follow its predicted paths. As shown in the acceleration analysis, the baseline model's jitter translates to extreme steering and braking maneuvers that exceed real-world vehicle limits.
The Pareto Frontier
PITA effectively navigates the "Pareto Frontier" of trajectory modeling. While the Simple Autoencoder achieves the lowest RMSE, it fails on physical metrics. Conversely, the Action-Space Autoencoder is perfectly physical but poor at reconstruction. PITA (specifically the $PITA_{Rec}$ variant) achieves nearly the same accuracy as the baseline while maintaining physical plausibility, making it the ideal "middle ground" for simulation.
Conclusion and Future Directions
PITA demonstrates that embedding physics as a "soft" constraint via regularization is a powerful method for generating high-fidelity trajectory data. While the current work focuses on single-agent dynamics in isolation, future iterations aim to incorporate map context (lane boundaries) and multi-agent interactions. Additionally, the authors plan to transition PITA from a deterministic autoencoder to a full generative framework, such as a Variational Autoencoder (VAE) or Diffusion Model, to facilitate the proactive discovery of novel safety-critical edge cases.
Summary updated