Qichao Zhang, Yinfeng Gao, Yikang Zhang, Youtian Guo, Dawei Ding, Yunpeng Wang, Peng Sun, Dongbin Zhao
Realistic and diverse simulation scenarios with reactive and feasible agent behaviors can be used for validation and verification of self-driving system performance without relying on expensive and time-consuming real-world testing. Existing simulators rely on heuristic-based behavior models for background vehicles, which cannot capture the complex interactive behaviors in real-world scenarios. To bridge the gap between simulation and the real world, we propose TrajGen, a two-stage trajectory generation framework, which can capture more realistic behaviors directly from human demonstration. In particular, TrajGen consists of the multi-modal trajectory prediction stage and the reinforcement learning-based trajectory modification stage. In the first stage, we propose a novel auxiliary RouteLoss for the trajectory prediction model to generate multi-modal diverse trajectories in the drivable area. In the second stage, reinforcement learning is used to track the predicted trajectories while avoiding collisions, which can improve the feasibility of generated trajectories. In addition, we develop a data-driven simulator I-Sim that can be used to train reinforcement learning models in parallel based on naturalistic driving data. The vehicle model in I-Sim can guarantee that the generated trajectories by TrajGen satisfy vehicle kinematic constraints. Finally, we give comprehensive metrics to evaluate generated trajectories for simulation scenarios, which shows that TrajGen outperforms either trajectory prediction or inverse reinforcement learning in terms of fidelity, reactivity, feasibility, and diversity.
Road testing for autonomous driving is expensive and time-consuming. Simulation has been an important evaluation tool enabling autonomous vehicles’ rapid development and safe deployment. The key is to construct realistic and diverse simulation scenarios, especially the behaviors of dynamic agents.
Recently, some trajectory prediction models (TrafficSim CVPR21'; SimNet ICRA21') are used as the behavior model of traffic agents. On one hand, much less attention has been paid to the feasibility of learning-based trajectory prediction, as those approaches model traffic agents as the particle model and generate future trajectories without imposing agent inherent kinematic constraints. On the other hand, as the total prediction steps go up, the accumulation of prediction errors will lead to many unreasonable behaviors such as collisions and off-road behaviors.
In this paper, we have proposed a novel two-stage trajectory generation framework to generate realistic and diverse trajectories with a reactive and feasible agent behavior model based on naturalistic driving data. The proposed TrajGen is given in the following figure. As shown in sub-figure(b), the trajectories of blue vehicles are generated by the trajectory prediction model in stage 1. Then, the vehicles with kinematic violation (agent 43) or collisions (agent 57) in stage 1 are controlled by the RL model to modify their predicted trajectories in stage 2, while the others are still following their predicted trajectories. For the proposed TrajGen, the fidelity and diversity of trajectories are provided by the trajectory prediction, and the reactivity and feasibility of trajectories are improved by RL. Finally, some realistic and diverse scenarios can be generated in I-Sim as shown in sub-figure(c).
Tracking consistency of agent behavior (red vehicle) based on this given bicycle model compared with log trajectory (white ghost vehicle which almost overlaps with the red vehicle).