Shuning1997/AGC-VLN-Town10HD-100episodes

AGC-VLN — Town10HD 100-Episode Results

The 100 closed-loop evaluation runs of AGC-VLN (Air-Ground Collaborative Vision-and-Language Navigation via Shared Bird's-Eye Maps) on the CARLA-Air Town10HD scene.

Overview

• 100 episodes = 50 scenes × 2 runs each.
• Aggregated metrics: joint success 77.0%, UAV success 50.0%, UGV success 75.0%, collaboration gain +27.0% (CG = SR_joint − min(SR_uav, SR_ugv)).
• Success threshold: either agent within 5 m of the goal.
• Each run records the final result.json plus per-step annotated bird's-eye maps, raw views, and VLM request/response JSON.

Data structure

• results.parquet — one row per run (100 rows) with the key metrics (loadable via datasets.load_dataset).
• results.csv — the same tabular metrics in CSV.
• summary.txt — aggregated batch summary.
• collab_* — the 100 raw run directories, each containing:
  • result.json — final metrics for the run (success flags, times, path lengths, final distances, VLM call counts).
  • step*_annotated.jpg / step*_birdview.jpg / step*_input.jpg / step*_fpv.jpg / step*_sim.jpg / step*_feedback.jpg — per-step visualizations.
  • step*_uav_vlm.json / step*_ugv_vlm.json — VLM request/response JSON.
  • log.txt — run log.

Parquet columns

Column	Description
ep	Episode id, e.g. town10hd_001
run	Run directory name
uav_success	UAV reached the goal within 5 m
ugv_success	UGV reached the goal within 5 m
joint	uav_success OR ugv_success (union)
time_s	Total episode time (s)
success_time_s	Time of the first arrival (s)
uav_success_time_s / ugv_success_time_s	Per-agent arrival time (s)
uav_path_m / ugv_path_m	Distance traveled (m)
uav_final_dist_m / ugv_final_dist_m	Final distance to goal (m)
uav_vlm_calls / ugv_vlm_calls / total_vlm_calls	VLM inference counts

Loading

from datasets import load_dataset

ds = load_dataset("Shuning1997/AGC-VLN-Town10HD-100episodes")
print(ds["train"].to_pandas())

Citation

If you use this dataset, please cite:

@article{zhang2026agcvln,
  title   = {Air-Ground Collaborative Vision-and-Language Navigation via Shared Bird's-Eye Maps},
  author  = {Shuning Zhang and Liang Li and Yunheng Wang and Tao Wang and Yihang Kang and Renjing Xu},
  year    = {2026},
  eprint  = {2609.03483},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
}