Evaluating Autonomous Vehicles’ Safety Benefits in Mixed Autonomy Scenarios

Joe-Wong, Carlee; Yağan, Osman; Lin, I-Cheng · 2024 · ROSA P / Carnegie Mellon University. Traffic21 Institute. Safety21 University Transportation Center (UTC)

archive: archived pipeline: cataloged verified

Get this paper ↗ (full text — opens at the source; we link to it, we don't host it)

Summary

This report evaluates the safety benefits of Connected Autonomous Vehicles (CAVs) in mixed-autonomy scenarios where CAVs share roads with human-driven vehicles, bicyclists, and pedestrians. The research addresses the uncertainty surrounding CAV safety impacts, noting that while CAVs promise to reduce human-error crashes, they may introduce new risks through interactions with humans or by encouraging risky behavior among human drivers. The study aims to understand both direct and indirect safety effects, such as changes in traffic density and flow smoothness, to determine the net impact of CAV deployment on vehicular safety. The methodology employs a simulation-based approach using a hybrid, hierarchical model that divides road networks into smaller "cells." This structure allows for microscopic modeling of individual vehicle interactions within cells, which are then summarized into cell-level statistics to determine broader network dynamics. The researchers utilized reinforcement learning simulations to model vehicle states, actions, and rewards, enabling the definition of various safety objectives and the incorporation of pedestrian presence. The simulator was extended from previous work to allow for arbitrary CAV objectives, discretized actions for both CAVs and human drivers, and flexible input for vehicle incident statistics, addressing the lack of reliable real-world CAV accident data. Key findings indicate that the simulator successfully reproduces known traffic phenomena, such as Braess’s paradox, where myopic individual travel time minimization leads to increased overall travel times. The study found that introducing CAVs programmed to minimize collective travel times can resolve this paradox, provided a sufficient proportion of vehicles are autonomous. However, the research highlights significant challenges in modeling safety dynamics. Reinforcement learning models struggled to accurately capture complex vehicle behavior and traffic dynamics due to the difficulty of training such models. Furthermore, the scarcity of real-world CAV incident data and the heavy dependence of traffic dynamics on specific road topologies limit the generalizability of findings. The study concludes that safety dynamics in mixed-autonomy environments are complex and difficult to predict, necessitating sophisticated, flexible simulators capable of modeling diverse CAV and human behaviors. The authors emphasize that developing realistic, safety-aware simulators requires more training data, which is currently difficult to obtain due to low CAV penetration rates. Future work should focus on characterizing road network types with similar safety and traffic characteristics to improve the generalizability of mixed-autonomy safety analyses.

Key finding

Safety dynamics in mixed-autonomy scenarios are complex and difficult to predict, requiring sophisticated simulators that can flexibly model a range of CAV and human driver behaviors, though training such models is hindered by limited real-world data.

Methodology

simulator

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via bulk_ingest_rosap on 2026-05-23 (6 acquisition events logged).

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• situational awareness

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Theoretical Contribution: computational model