S4TP: Social-Suitable and Safety-Sensitive Trajectory Planning for Autonomous Vehicles

TL;DR

S4TP tackles Safe trajectory planning for autonomous vehicles in socially interactive mixed traffic by combining Transformer-based Social-Aware Trajectory Prediction (SATP) with a Social-Aware Driving Risk Field (SADRF). SATP predicts HDV trajectories through scene encoding that captures agent-map and agent-agent interactions, while SADRF builds a dynamic 2D risk heat map around the AV to guide online optimization-based planning within low-risk regions. The framework is validated in SMARTS with four challenging scenarios, showing superior safety, efficiency, and human-like behavior, including a 100% pass rate across scenarios and clear gains over state-of-the-art baselines. This approach enhances interpretability and practicality for real-world autonomous driving in social traffic, enabling safer and more natural AV behavior in complex environments.

Abstract

In public roads, autonomous vehicles (AVs) face the challenge of frequent interactions with human-driven vehicles (HDVs), which render uncertain driving behavior due to varying social characteristics among humans. To effectively assess the risks prevailing in the vicinity of AVs in social interactive traffic scenarios and achieve safe autonomous driving, this article proposes a social-suitable and safety-sensitive trajectory planning (S4TP) framework. Specifically, S4TP integrates the Social-Aware Trajectory Prediction (SATP) and Social-Aware Driving Risk Field (SADRF) modules. SATP utilizes Transformers to effectively encode the driving scene and incorporates an AV's planned trajectory during the prediction decoding process. SADRF assesses the expected surrounding risk degrees during AVs-HDVs interactions, each with different social characteristics, visualized as two-dimensional heat maps centered on the AV. SADRF models the driving intentions of the surrounding HDVs and predicts trajectories based on the representation of vehicular interactions. S4TP employs an optimization-based approach for motion planning, utilizing the predicted HDVs'trajectories as input. With the integration of SADRF, S4TP executes real-time online optimization of the planned trajectory of AV within lowrisk regions, thus improving the safety and the interpretability of the planned trajectory. We have conducted comprehensive tests of the proposed method using the SMARTS simulator. Experimental results in complex social scenarios, such as unprotected left turn intersections, merging, cruising, and overtaking, validate the superiority of our proposed S4TP in terms of safety and rationality. S4TP achieves a pass rate of 100% across all scenarios, surpassing the current state-of-the-art methods Fanta of 98.25% and Predictive-Decision of 94.75%.

Figures (4)

• Figure 1: The overall architecture of the proposed S$^{\text{4}}$TP.
• Figure 2: Illustration of modeling the SADRF. To implement our proposed SADRF modeling, we begin by calculating the predicted trajectories of HDVs based on the vehicle dynamics model, as depicted in (a), where we use the state inputs as the position($x_{\text{car}},y_{\text{car}}$), heading $\phi_{\text{car}}$ and the steering angle $\delta$ of HDVs. The predicted path (arc length $s$) is calculated by $s =v*t_{\text{la}}$. (b) Showcases the dynamic generation of DRF, specifically during the planning of an unprotected left-turn path. (b1), (b2), and (b3) illustrate path planning process, visualization of SADRF alone, and visualization of SADRF incorporated into the scene, respectively. The individual SADRF graph on the right side represents the risk values within the scene using a color-coded bar chart ranging from light to dark red. In this visualization, the highest risk value is assigned to HDVs. The risk value is dynamically computed based on the currently planned trajectory, adapting to changes at subsequent moments during dynamic planning. (c) Demonstrate the shape of the SADRF with respect to velocity $v$ and steering angle $\delta$. As can be seen, the SADRF expands as the velocity increases and the steering angle increases, thus improving the security for subsequent trajectory planning.
• Figure 3: Four designed driving scenarios in the SMARTS environment.
• Figure 4: Visual comparison of planning trajectories between our proposed S$^{\text{4}}$TP, the baseline Prediction-Decision, and the static DRF in unprotected left-turn intersection scenario. The red box and line are the AV and its planned trajectory; the blue boxes are the surrounding agents. (a) The proposed SADRF-based planner excels at accurately assessing surrounding risk levels in complex social interaction scenarios, thus ensuring the safety of the AV. In contrast, (b) the Prediction-Decision planner, which may disregard risk fields, fails to accurately assess the surrounding risk levels and consequently results in collisions. (c) The static DRF, lacking the ability to dynamically assess trajectory risk values, leads to overly conservative strategy generation in complex scenarios compared to SADRF.