Investigating Driving Interactions: A Robust Multi-Agent Simulation Framework for Autonomous Vehicles

TL;DR

The paper tackles the challenge of validating autonomous vehicles in complex, interactive traffic by introducing a synchronous multi-agent simulation framework that replaces non-reactive traffic with intelligent agents and plugs in diverse trajectory planners via a CommonRoad extension. It formalizes the simulation with a time-discrete update (Δt = 0.1s), a 3 s motion-prediction horizon, local perceptions, and a planner interface, with open-source implementation. It contributes a formal model, a benchmarking suite for agent performance and behavioral safety in curvilinear coordinates, runtime analyses, and demonstrations in lane merging and intersection tasks, using FRENETIX as a benchmarking planner and WALENET for motion prediction. The framework enables deterministic, reproducible experimentation for developing robust planning algorithms and V2V research in complex driving scenarios, with potential extensions to cooperative planning and reinforcement learning.

Abstract

Current validation methods often rely on recorded data and basic functional checks, which may not be sufficient to encompass the scenarios an autonomous vehicle might encounter. In addition, there is a growing need for complex scenarios with changing vehicle interactions for comprehensive validation. This work introduces a novel synchronous multi-agent simulation framework for autonomous vehicles in interactive scenarios. Our approach creates an interactive scenario and incorporates publicly available edge-case scenarios wherein simulated vehicles are replaced by agents navigating to predefined destinations. We provide a platform that enables the integration of different autonomous driving planning methodologies and includes a set of evaluation metrics to assess autonomous driving behavior. Our study explores different planning setups and adjusts simulation complexity to test the framework's adaptability and performance. Results highlight the critical role of simulating vehicle interactions to enhance autonomous driving systems. Our setup offers unique insights for developing advanced algorithms for complex driving tasks to accelerate future investigations and developments in this field. The multi-agent simulation framework is available as open-source software: https://github.com/TUM-AVS/Frenetix-Motion-Planner

Figures (8)

• Figure 1: Illustrative example of the multi-agent simulation framework with three agents, each controlled by a sampling-based motion planner.
• Figure 2: The multi-agent simulation framework structure illustrated with a scenario involving four agents: In each time step, the states of all agents are aggregated in the global scenario, where a collision check is conducted. Subsequently, the future movement of all vehicles is predicted, and the information is shared with each agent. Every agent maintains a local representation of the scenario, including only the positions and predictions of surrounding vehicles. Based on this information, agents calculate their next trajectory step. The yellow and orange areas in the figures of each agent show the probability-based motion predictions, while the spawned trajectories are the sampled paths to select the next motion step. In the global scenario, every agent follows their individual best trajectory. When the simulation is finished, the final trajectories are evaluated with criticality measures.
• Figure 3: Difference between curvilinear and Cartesian distances. Green: Reference paths of the two curvilinear coordinate systems at the junction spanned for the orange vehicle.
• Figure 4: Average computation time per simulation step with single-core and multi-core computation (8 cores). Error bars indicate the first and third quartile of the logged computation time.
• Figure 5: Lane merging scenario: Comparison of the driving interactions. The orange vehicle is an intelligent agent, while the interaction abilities of the green car are altered in each simulation.