Benchmark and Baselines for Desired-Speed Conditioned Autonomous Driving
📄 arXiv | 🌐 Website | 📁 Dataset | 🤖 Baseline
Bench2Drive-Speed is a closed-loop benchmark for desired-speed conditioned autonomous driving, enabling explicit control over vehicle behavior through target speed and overtake/follow commands.
End-to-end autonomous driving (E2E-AD) systems have achieved strong performance, but lack explicit controllability over user preferences such as:
- Desired driving speed
- Whether to overtake or follow other vehicles
Bench2Drive-Speed fills this gap by introducing:
- 🎯 Target-speed conditioning
- 🚘 Overtake / Follow commands
- 📊 Quantitative controllability metrics
- 🔁 Closed-loop evaluation benchmark
Bench2Drive-Speed's CustomizedSpeedDataset contains 2,100 CARLA driving scenarios with expert demonstrations and annotated overtake/follow commands.
The released dataset includes expert target-speed signals only.
🔗 Download:
- 🤗 Huggingface: https://huggingface.co/datasets/rethinklab/Bench2Drive-Speed
If you downloaded this, don't forget to extract the tar.gz archive.
You can use a script to generate virtual target speed from Bench2Drive-style driving dataset.
-
Find
tools/append_virtual_target_speed.py -
Set the path to your dataset in the main function.
# ====================== # entry # ====================== if __name__ == "__main__": main( ["/path/to/dataset"] # << modify here )
-
Set hyperparameters. In our paper's setting,
-
Short:
MAX_EXTEND = 3.0 TIME_MAX = 1.5 -
Long:
MAX_EXTEND = 10.0 TIME_MAX = 3.0
-
Run the script.
python tools/append_virtual_target_speed.py
For carla setup, please refer to Bench2Drive repo.
To create a new agent, please refer to tcp_b2d_speed_agent.py, which implements an agent that inherits from autonomous_agent.AutonomousAgent.
In addition to the default CARLA autonomous driving agents, you can use self.get_planned_speed() to obtain the current target speed as a float (km/h). You can also use self.do_overtake to access the current overtaking command as a boolean.
-
Add your agent to leaderboard/team_code/your_agent.py & Link your model folder under the Bench2Drive directory.
Bench2Drive\ assets\ docs\ leaderboard\ team_code\ --> Please add your agent HERE scenario_runner\ tools\ --> Please link your model folder HERE
-
To do the evaluation, please write a start up script like this under
Bench2Drive/leaderboard/scripts(this one is wrote for TCP-Speed):#!/bin/bash BASE_PORT=30000 # CARLA Simulator's port BASE_TM_PORT=50000 # Traffic manager's port IS_BENCH2DRIVE=True # Please set this to true BASE_ROUTES=leaderboard/data/b2dspd_eval48 # Path to the evaluation route json TEAM_AGENT=leaderboard/team_code/your_agent.py # Path to your agent TEAM_CONFIG=your_team_agent_ckpt.ckpt # Path to your config. # Agents like uniad/vad's configuration might different from this one # please refer to Bench2Drive repo. BASE_CHECKPOINT_ENDPOINT=checkpoint_name # Path to the json file which records evaluation result and progress SAVE_PATH=output # The directory where sensor output and data used for metrics are saved PLANNER_TYPE=only_traj # Only for TCP, you can delete this line GPU_RANK=2 # The gpu you want CARLA to run on. # Notably, CARLA ranks gpu according to vulkan, # so it might not match the index in nvidia-smi. CUDA_VISIBLE_DEVICES=2 # The gpu you want your model to run on. PORT=$BASE_PORT TM_PORT=$BASE_TM_PORT ROUTES="${BASE_ROUTES}.xml" CHECKPOINT_ENDPOINT="${BASE_CHECKPOINT_ENDPOINT}.json" export CLEAR_LANE=1 # Make sure this environ is set! bash leaderboard/scripts/run_evaluation.sh $PORT $TM_PORT $IS_BENCH2DRIVE $ROUTES $TEAM_AGENT $TEAM_CONFIG $CHECKPOINT_ENDPOINT $SAVE_PATH $PLANNER_TYPE $GPU_RANK
-
Start the evaluation:
cd Bench2Drive/ # Verify the correctness of the team agent, need to set GPU_RANK, TEAM_AGENT, TEAM_CONFIG bash leaderboard/scripts/run_evaluation_tcp_speed.sh # or execute your custom startup script if you have defined one.
You can modify this script to adapt your agent.
-
Metric: Make sure there are exactly 48 routes in your json. Failed/Crashed status is also acceptable. Otherwise, the metric is inaccurate.
When using the scripts below, remember to modify the directory parameters to yours.
# Get Speed-Adherence python tools/speed_metric_new.py # This script will generate speed adherence score and visualized speed curves. # Get Driving Score and Success Rate python tools/metric_stats_48.py # This script will generate three files: # 1. {output_name}_with_overtake.json, # in which the overtake/follow infraction is recorded # and calculated as x0.7 penalty; # 2. {output_name}_traditional.json, # in which the overtake/follow infraction does not count; # 3. {output_name}_statistic.json, # in which only evaluation results are recorded, # excluding detailed infractions # Get Overtake Score python tools/overtake_score.py # This script will generate overtake score # Get multi-ability results python tools/ability_benchmark_offline.py -r {output_name}_traditional.json # Get driving efficiency and driving smoothness results python tools/efficiency_smoothness_benchmark.py -f {output_name}_traditional.json -m your_metric_folder/
🔗 You can download baseline checkpoints to avoid training from scratch:
- 🤗 Huggingface: https://huggingface.co/rethinklab/TCP-Speed
-
Preprocess data
cd TCP # need set YOUR_Data_PATH python tools/gen_tcp_data_local.py # this script will generate dataset with target speed and virtual target speed
-
Training: using expert demonstration
cd TCP bash TCP/train_speed.sh -
Training: using virtual target speed (make sure they are annotated! see Virtual Target Speed Annotation)
cd TCP bash TCP/train_speed_virtual.sh
All assets and code are under the CC-BY-NC-ND unless specified otherwise.
@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},
}