VAD-Speed
VAD-Speed is a baseline model for desired-speed conditioned autonomous driving, introduced as part of the Bench2Drive-Speed benchmark. For other baseline models, you can find TCP-Speed here.
Project Page | Paper | Github
Overview
Bench2Drive-Speed is a closed-loop benchmark for desired-speed conditioned autonomous driving. It enables explicit control over vehicle behavior through:
- 🎯 Target-speed conditioning: Users can specify the desired driving speed.
- 🚘 Overtake / Follow commands: Users can specify whether the vehicle should overtake or follow others.
The VAD-Speed model serves as a baseline that processes these user preferences to generate driving actions within the CARLA simulator.
Usage
To use or build upon this agent, refer to the official Bench2Drive-Speed repository.
The agent inherits from AutonomousAgent. In addition to standard CARLA inputs, you can access the user-specified conditions within the agent logic as follows:
# Extract target speed and overtake commands within the agent class
target_speed = self.get_planned_speed() # float (km/h)
do_overtake = self.do_overtake # boolean
For full setup, training, and evaluation instructions, please follow the documentation in the Github README.
Citation
@article{Bench2DriveSpeed,
title={Can Users Specify Driving Speed? Bench2Drive-Speed: Benchmark and Baselines for Desired-Speed Conditioned Autonomous Driving},
author={Yuqian Shao and Xiaosong Jia and Langechuan Liu and Junchi Yan},
year={2026},
eprint={2603.25672},
archivePrefix={arXiv},
primaryClass={cs.RO},
}