Research on Transformer Models for End-to-End Control in Autonomous Driving

Engineering notes

Experimental results show that this paper's approach outperforms the benchmark CNN-LSTM and PilotNet models, achieving performance improvements of 44% in control accuracy (MSE) and 63.5% in driving safety. Additionally, the proposed model demonstrates superior performance in control accuracy, driving safety, real-time response, robustness, and interpretability.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

This study addresses the challenge of trajectory tracking control in autonomous vehicles. Traditional hierarchical control methods often require manual parameter tuning and struggle to adapt to complex, multi-modal environments. To overcome these limitations, this paper proposes a Transformer-based end-to-end control model for autonomous driving. The model leverages self-attention mechanisms to dynamically fuse multi-modal inputs and capture long-term temporal dependencies. It consists of three main components: input encoding, multi-modal feature fusion, and control signal decoding. This paper evaluates the proposed model using datasets collected from the CARLA simulator and trains it with a hybrid training strategy training strategy. Experimental results show that this paper's approach outperforms the benchmark CNN-LSTM and PilotNet models, achieving performance improvements of 44% in control accuracy (MSE) and 63.5% in driving safety. Additionally, the proposed model demonstrates superior performance in control accuracy, driving safety, real-time response, robustness, and interpretability. Further analysis shows that incorporating multi-frame temporal inputs, an 8-head attention mechanism, and a cross-attention fusion strategy enhances model performance, highlighting its strong potential for real-world applications.

Links and sources

• Official / arXiv page

Comments