Efficient End-to-End Deep Learning for Autonomous Racing: TinyLidarNet and Low-Power Computing Platforms

End-to-end deep learning has proven effective for robotic navigation by deriving control signals directly from raw sensory data. However, the majority of existing end-to-end navigation solutions are predominantly camera-based. 

We propose TinyLidarNet, a lightweight 2D LiDAR-based end-to-end deep learning model for autonomous racing. We systematically analyze its performance on untrained tracks and computing requirements for real-time processing. We find that TinyLidarNet's 1D Convolutional Neural Network (CNN) based architecture significantly outperforms widely used Multi-Layer Perceptron (MLP) based architecture. In addition, we show that it can be processed in real-time on low-end micro-controller units (MCUs).

We deployed TinyLidarNet on an MCU-based F1TENTH platform, which is comprised of an ESP32-S3 MCU and a RPLiDAR sensor and demonstrated the feasibility of using MCUs in F1TENTH autonomous racing. 

Finally, we compare TinyLidarNet with ForzaETH, a state-of-the-art Model Predictive Controller (MPC) based F1TENTH racing stack. Our results show that TinyLidarNet is able to closely match the performance of ForzaETH by training the model using the data generated by ForzaETH