Training and Evaluation on Bench2Drive

Training

Stage 1

cd Bench2Drive
bash tools/dist_train.sh projects/configs/unidrivevla_b2d_stage1_unified_2b.py 8

Stage 2

cd Bench2Drive
bash tools/dist_train.sh projects/configs/unidrivevla_b2d_stage2_unified_2b.py 8

---

Closed-Loop Evaluation

1. Create Multi-Route Splits

python ./bench2drive/tools/split_xml.py

2. Configure Paths

Set WORK_DIR and CARLA_ROOT in the evaluation script:

# bench2drive/leaderboard/scripts/run_evaluation.sh
export WORK_DIR=/path/to/unidrivevla/Bench2Drive
export CARLA_ROOT=/path/to/unidrivevla/Bench2Drive/carla

3. Run Evaluation

Before running, set the following variables in bench2drive/leaderboard/scripts/run_evaluation_multi_unidrivevla.sh, or pass them as environment variables:

Variable	Description
WORK_DIR	Absolute path to the Bench2Drive/ directory
CHECKPOINT	Path to the model weights (.pt file)
SAVE_PATH	Output directory for evaluation results (default: evaluation/unidrivevla_b2d)

cd Bench2Drive
WORK_DIR=/path/to/unidrivevla/Bench2Drive \
CHECKPOINT=/path/to/UniDriveVLA_Stage2_Bench2Drive_2B.pt \
SAVE_PATH=evaluation/unidrivevla_b2d \
bash bench2drive/leaderboard/scripts/run_evaluation_multi_unidrivevla.sh

4. Compute Metrics

python ./bench2drive/tools/statistic_route_json.py \
    --route_dir evaluation/unidrivevla_b2d

5. Generate Videos (Optional)

python ./bench2drive/tools/generate_video.py \
    --folder_dir evaluation/unidrivevla_b2d