Introduction
Testing advanced control and state estimation algorithms on physical hardware is a critical step in robotics research, yet it remains a significant bottleneck due to the high costs of infrastructure. High-fidelity motion capture systems, such as Vicon or Qualisys, often cost tens of thousands of dollars and require dedicated laboratory space. This financial barrier frequently limits researchers to simulation, leading to the "sim-to-real" gap where algorithms fail to translate to the complexities of the physical world.
This paper introduces a fully open-source, low-cost (under $700) miniature car-like robotics platform designed to democratize high-performance autonomous racing research. By leveraging a sophisticated optimization-based software pipeline, the authors demonstrate that expensive motion capture systems are no longer a prerequisite for accurate state estimation and robust control.
System Modeling and Identification
The platform is built upon the "Chronos" miniature RC car, modified with a custom ESP32-based PCB, an IMU, magnetic wheel encoders, and a "Lighthouse" positioning deck. To achieve high-performance control on such hardware, the authors developed a specialized dynamic model and identification process.
Modified Dynamic Bicycle Model
The researchers utilized a bicycle model incorporating Pacejka tire forces. To adapt this model for a wider range of maneuvers, they introduced two significant modifications:
- AWD Force Distribution: A parameter was added to split motor force for All-Wheel Drive (AWD) configurations.
- Singularity Mitigation: Traditional Pacejka models suffer from mathematical singularities at low velocities. The authors addressed this by splicing a 3rd-degree polynomial fit for low-speed regimes, ensuring the model remains continuously differentiable and numerically stable from a standstill to high speeds.
Optimization-Based System ID
To calibrate the model, the authors employed a nonlinear programming (NLP) approach for offline system identification. By processing driving data, the NLP solver automatically identifies physical parameters that are difficult to measure directly, such as tire friction coefficients and vehicle inertia. This process significantly improved model fidelity, reducing open-loop prediction Root Mean Square Error (RMSE) from 0.24 meters to 0.09 meters over a 2-second horizon.
State Estimation and Control
Moving Horizon Estimation (MHE)
Rather than relying on a standard Extended Kalman Filter (EKF), the system utilizes Moving Horizon Estimation (MHE) for real-time state tracking. MHE treats estimation as an optimization problem over a sliding window of past measurements. This approach is inherently more robust to sensor noise and intermittent data dropouts than recursive filters.
Control Architecture
The state estimates from the MHE are fed into a Model Predictive Contouring Controller (MPCC). The MPCC is tasked with navigating the car around a track at the limits of handling while respecting the physical constraints of the vehicle and the environment.
Experimental Results and Dataset
Localization and Estimation Performance
The low-cost Lighthouse positioning system, which utilizes HTC Vive base station technology, achieved remarkable precision with a mean residual error of just 3 mm after a simple calibration using 5–6 static points. During active driving, the MHE maintained a position RMSE of approximately 32 mm on the x-axis and 21 mm on the y-axis.
Robustness to Sensor Failure
A key highlight of the study was the system's resilience. During hardware experiments, the authors intentionally disabled the Lighthouse positioning system for 1.127 seconds (equivalent to roughly 68 missed MHE updates). The MHE successfully maintained stable state estimates by fusing data from the IMU and wheel encoders, allowing the MPCC to continue navigating the track without a collision.
Open-Source Contribution
To support the community, the authors have released a comprehensive real-world dataset. This dataset includes synchronized data from the IMU, wheel encoders, and Lighthouse system, alongside high-fidelity ground truth from a Qualisys motion capture system. This serves as a benchmark for future research in nonlinear system identification and state estimation.
Discussion and Future Work
While the optimization-based pipeline proved highly effective, the authors noted that integrating the low-cost IMU occasionally degraded velocity estimates due to high noise levels and bias drift. Furthermore, while MHE provided superior robustness during sensor dropouts, its baseline accuracy was comparable to an EKF under normal conditions. Future research will focus on identifying specific highly-nonlinear regimes where MHE provides a definitive performance advantage and improving the integration of noisy inertial data.
Summary updated