A Low-Cost, Open Platform for End-to-End Autonomous Driving on a Miniature Ackermann Vehicle
A paper presents a low-cost, open experimental platform for end-to-end autonomous driving research using miniature Ackermann vehicles. The platform includes a physical vehicle, printed urban track, data collection tools, trajectory registration, and a Webots digital twin. A command-conditioned behavior cloning baseline uses an on-board camera image and high-level navigation command to output steering and speed. In real closed-loop experiments, the learned policy achieved a mean cross-track error of 6.1 cm versus 4.7 cm for human demonstrations. In the digital twin, widening camera field of view from 58 to 120 degrees reduced mean cross-track error from 35.6 to 3.3 cm. The paper was published on arXiv on 2026-09-03.
Researchers introduced a low-cost, open platform for end-to-end autonomous driving on miniature Ackermann vehicles, combining physical hardware, a printed urban track, data collection, trajectory registration, and a Webots digital twin. A baseline command-conditioned behavior cloning policy takes camera images and navigation commands to output steering and speed. Real-world tests showed a mean cross-track error of 6.1 cm, close to the 4.7 cm human baseline. Simulation experiments revealed that a wider camera field of view (120° vs 58°) dramatically improved performance, reducing error from 35.6 cm to 3.3 cm. The platform enables controlled experiments linking simulation and real-world autonomous driving.
The platform integrates a physical miniature Ackermann vehicle with a Webots digital twin, enabling sim-to-real transfer experiments. The behavior cloning baseline uses a neural policy with camera image and navigation command inputs. Key finding: camera field of view is a critical factor; widening from 58° to 120° reduced cross-track error by over 90% in simulation. The paper also mentions using synthetic driving data and a learned sim-to-real image translator to reduce appearance gap, suggesting a domain adaptation approach. The low-cost, open nature makes it accessible for reproducible research.
This work lowers the barrier for academic and small-lab research in end-to-end autonomous driving by providing an affordable, open-source hardware and software stack. It highlights the importance of sensor configuration (field of view) for learned driving policies, which could inform design choices in larger-scale autonomous systems. The use of digital twins for data generation and sim-to-real transfer is a growing trend in robotics and autonomous vehicle development, potentially reducing reliance on expensive real-world data collection.
The open platform could accelerate research and development in autonomous driving by reducing costs and enabling rapid prototyping. It may attract interest from educational institutions, startups, and companies exploring miniature vehicle testbeds. The findings on camera field of view could influence sensor selection in commercial autonomous systems, potentially leading to cost savings or performance improvements. However, direct commercial impact is limited at this stage as it is a research platform.
Next signals to watch include: release of the platform's code and hardware specifications, adoption by other research groups, extensions to more complex navigation tasks or multi-vehicle scenarios, and improvements in sim-to-real transfer techniques. The platform could serve as a benchmark for comparing end-to-end driving algorithms. If the sim-to-real translator proves effective, it may encourage more use of synthetic data in autonomous driving research.