Long, Tingting (Xi'an Jiaotong University), Zong, Mengru (Xi'an Jiaotong University), Zhang, Chi (Xi'an Jiaotong University), Liu, Yuehu (Institute of Artificial Intelligence and Robotics, Xi'an Jiaoton), Ma, Shuangxun (Chang'an University), Li, Li (Tsinghua University)

Generating Realistic VRU-AV Scenarios Via Social Force-Based Gradient Optimization

Scheduled for presentation during the Regular Session "Generating driving scenarios II" (FrBT10), Friday, September 27, 2024, 15:10−15:30, Salon 18

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on October 8, 2024

Abstract

Validating safe and reliable interaction with vulnerable road users (VRUs) is imperative for autonomous vehicles, especially in simulation testing procedures before on-road applications. However, existing VRU agents are either with predefined routes or with collision-oriented behaviors, which limits the fidelity and diversity of scenarios. Bridging the gap between the simulated world and reality requires realistic VRU behaviors. In this paper, we apply social force to capture natural human behavior, considering how individuals are influenced by surroundings and adhere to real-world rules. To this end, we propose HybridSFM, an improved social force model designed for characterizing human-vehicle interacting behavior. Specifically, the steering and acceleration of the VRUs are parameterized and integrated with social norms to forecast trajectories. Based on this, we further implement a gradient optimization framework for generating VRU-AV safety-critical scenarios. Firstly, VRUs are positioned randomly around the ego-agent. Differentiable simulation updates states and unrolls trajectories while backpropagation optimizes the actions of VRUs. The modified scenario is further optimized in the subsequent timestep until a collision happens. The experiments demonstrate that our approach can enhance realistic safety-critical scenarios at intersection scenarios, facilitating the testing of interaction capabilities for autonomous vehicles.