An End-to-End Autonomous Driving Pre-trained Transformer Model for Multi-Behavior-Optimal Trajectory Generation

Engineering notes

Key topics: end-to-end autonomous driving, autonomous driving, end-to-end, reinforcement learning, imitation learning, planning. See the paper for implementation details and experimental results.

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

Advanced end-to-end (E2E) autonomous driving planning methods, e.g., reinforcement learning and imitation learning, employ neural networks to generate driving trajectories. However, most of them suffer from low data efficiency, i.e, they usually require a large amount of data to obtain even a simple behavior like car-following which a rule-based method can easily achieve. To this end, this work tackles the problem by introducing a pre-training method for E2E planning, which can generate multiple initial near-optimal trajectories for further fine-tuning with specific datasets. The challenges lie in that 1) various driving behaviors may be involved in a single case, leading to different optimal trajectories; 2) abstract driving behaviors are hard to be manually defined. Our idea is to first design a multi-head transformer network structure to capture the optimal trajectories for different driving behaviors. The training trajectory dataset is formed by an offline trajectory optimizer with different convex corridors which grow from sampled trajectories and imply various driving behaviors. In this way, the model can be pre-trained with various behavior-optimal trajectories, which can cover the whole driving space without manually defined behaviors. The results show that the pre-trained model can generate multiple trajectories for different behaviors without collected human demonstrations. This work can accelerate the future E2E model fine-tuning, with an acceptable initial baseline performance.

Links and sources

• Official / arXiv page

Comments