Abstract
Bridging the "sim-to-real" gap remains a critical challenge in autonomous vehicle robotics. When simulators fail to accurately reflect physical properties like mass, tire stiffness, and center of gravity, controllers designed in simulation often fail in real-world deployment. This paper presents a framework for System Identification (SysID) of front-steered Ackermann vehicles using a Differentiable Physics Simulator (DPS). By leveraging analytical gradients, our method identifies physical parameters more efficiently than traditional gradient-free optimization techniques. We validate this approach using real-world data from the F1TENTH platform, demonstrating that parameters learned through DPS enable stable lane-keeping control comparable to models using ground-truth specifications.
Introduction
System Identification (SysID) is the process of determining the physical parameters of a robotic system to ensure its mathematical model matches its real-world behavior. For vehicles following Ackermann geometry—such as cars, trucks, and tractors—accurate SysID is essential for high-performance maneuvering. Traditionally, this process involves manual measurement or expensive specialized equipment. While machine learning has been applied to SysID for skid-steer robots and robotic arms, front-steered vehicles have received less attention. This study introduces a gradient-based framework that utilizes differentiable physics to automate and optimize the identification of these critical vehicle dynamics.
Methodology
The proposed framework integrates a custom simulator with gradient-based optimization to refine vehicle parameters.
Differentiable Physics Simulator (DPS)
Built using the PyTorch library, the simulator provides analytical gradients. Unlike "black-box" simulators that only provide the final error, a DPS allows the system to calculate exactly how a specific parameter (e.g., tire friction) should be adjusted to minimize the difference between simulated and observed trajectories. This differentiability significantly accelerates the learning process.
Vehicle Dynamics Model
The study employs a standard 2-degree-of-freedom (2-DOF) bicycle model to represent lateral vehicle dynamics. This model focuses on the relationship between steering input and the resulting lateral forces and yaw rate, which are fundamental to Ackermann steering systems.
Hybrid Loss Function
To accurately compare simulated trajectories against real-world observations, the authors developed a hybrid loss function:
- Dynamic Time Warping (DTW): Used to align trajectories in time, ensuring that temporal variations do not skew the optimization.
- Chamfer Distance: Used to ensure the spatial scale of the trajectories matches. The authors found that DTW alone could lead the model to "hallucinate" correctly shaped but incorrectly sized trajectories; the Chamfer distance provides the necessary spatial constraint.
Optimization and Control
The gradient-based approach (using the Adam optimizer) was benchmarked against Covariance Matrix Adaptation Evolution Strategy (CMA-ES), a popular gradient-free baseline. Once parameters were identified, a state-feedback controller for lane keeping was designed using a pole placement algorithm.
Experimental Setup
Data was collected using an F1TENTH autonomous racing platform equipped with a high-precision Phasespace X2E LED motion capture system for ground-truth state estimation. The dataset consisted of 16 circular trajectories (8 left, 8 right) at a constant velocity of 1 m/s. To test the robustness of the SysID, experiments were conducted under two conditions: standard rubber tires and low-friction tires (wrapped in tape).
Results and Discussion
Efficiency and Convergence
The gradient-based method demonstrated superior sample efficiency compared to the CMA-ES baseline. Because the optimizer had access to analytical gradients, it converged to the correct physical parameters in significantly fewer iterations, exhibiting much smoother loss curves during the training process.
Parameter Accuracy
The learned parameters—including mass, vehicle length, and cornering stiffness—closely aligned with the official F1TENTH documentation. Interestingly, in low-friction tests, the identified cornering stiffness was slightly higher than theoretical expectations. This suggests that while the model is highly accurate, capturing the full complexity of tire-ground interactions may require training data that includes higher accelerations and steering velocities to push the tires to their physical limits.
Real-World Deployment
The ultimate test of the SysID framework was the performance of the resulting controller. In real-world lane-keeping trials, the controller utilizing the learned parameters performed almost identically to a controller designed with official ground-truth parameters. This confirms that the DPS-based identification is accurate enough for practical robotic applications.
Conclusion and Future Work
This research demonstrates that differentiable physics is a powerful tool for the system identification of Ackermann vehicles. By providing a path toward automated, efficient, and accurate parameter estimation, this framework helps close the sim-to-real gap for autonomous driving systems.
Future research will focus on transitioning toward online system identification, allowing robots to update their internal models in real-time as environmental conditions change. Additionally, the authors plan to explore more complex maneuvers involving varying speeds and high-dynamic accelerations to better capture non-linear tire dynamics.
Summary updated