Presents a low-cost miniature Ackermann-vehicle platform with physical track, Webots twin, and trajectory registration; demonstrates command-conditioned behavior cloning as baseline for sim-to-real autonomous driving.
Behavior Cloning for Mini Autonomous Car Path Following presents a low-cost, open miniature Ackermann vehicle platform for end-to-end autonomous driving, demonstrating command-conditioned behavior cloning as a baseline for sim-to-real transfer. The system integrates a physical track, a Webots digital twin, and trajectory registration to enable robust learning.
Key components of the platform include: Hardware: A Jetson Nano 4GB compute unit, Logitech C920 HD Pro camera, and 6V DC motors with an SG5010 servo for Ackermann steering. Software: Behavior Cloning (BC) maps camera data to control outputs using a Convolutional Neural Network (CNN), avoiding complex algorithmic pipelines. * Validation: The approach is scaled from mini-autonomous cars to real-world implementations, providing a foundation for more advanced autonomous navigation strategies.
The paper presents an open, low-cost testbed for end-to-end autonomous driving built around a miniature Ackermann-steering vehicle and a physical driving track. The platform is designed to make the full autonomous-driving loop—sensing, policy inference, actuation, and evaluation—accessible at a scale and cost that are practical for research labs, classrooms, and rapid prototyping. A central component is a Webots digital twin of the vehicle and environment, complemented by a trajectory-registration process that aligns real-world vehicle trajectories with their simulated counterparts. This registration step is important because it helps reduce the geometric and coordinate-system mismatches that often degrade sim-to-real transfer, allowing policies trained in simulation to be more meaningfully evaluated on the physical vehicle.
A key contribution is the demonstration of a simple but useful baseline: command-conditioned behavior cloning for sim-to-real driving. In this setup, a policy is trained in the Webots twin to map observations and high-level driving commands to low-level vehicle control actions, and the resulting policy is then deployed on the miniature Ackermann vehicle. The paper does not primarily claim a new state-of-the-art driving architecture; rather, it provides a reproducible reference implementation and a concrete demonstration that imitation-learning methods can transfer from a structured simulator to a small physical platform. This is valuable because it gives researchers a baseline against which more advanced end-to-end, reinforcement-learning, or perception-heavy approaches can be compared.
The work matters because it lowers the barrier to entry for studying autonomous driving while preserving many of the practical challenges of the full-scale problem: sensor noise, actuator response, Ackermann steering dynamics, tracking errors, and the sim-to-real gap. By combining an affordable hardware platform, an open simulation environment, and a trajectory-registration methodology, the paper supports more reproducible experimentation and standardized evaluation. For a technically literate audience, the contribution is less about a single algorithmic breakthrough and more about providing a practical, transparent, and extensible platform for investigating end-to-end driving policies in a controlled miniature setting.