WHALES

A Multi-Agent Scheduling Dataset for Enhanced Cooperation in Autonomous Driving

Introduction

WHALES (Wireless enHanced Autonomous vehicles with Large number of Engaged agentS) is a CARLA-based cooperative perception dataset averaging 8.4 agents per sequence. It captures diverse viewpoints, agent behaviors, and multitask interactions to study scheduling, perception, and planning under realistic multi-agent constraints.

News

• 2025-11-21 – Released WHALES dataset v1.0 with cooperative scheduling benchmarks.
• 2025-06-17 – WHALES was accepted by IROS 2025!

Table of Contents

• Dataset Overview
• Getting Started
• Training & Evaluation
• Visualization
• Scheduling Algorithms
• Experimental Results
• Roadmap
• Citation

Highlights

• Largest agent count: 8.4 agents per scene with synchronized LiDAR-camera suites.
• Rich annotations: 2.01M 3D boxes plus full agent-behavior recording.
• Scheduling-ready: Provides perception, planning, and communication metadata for agent selection research.
• Plug-in friendly: Ships with mmdetection3d-compatible configs and hooks for custom schedulers.

Dataset Overview

Comparison with Existing Benchmarks

Dataset	Year	Real/Simulated	V2X	Image	Point Cloud	3D Annotations	Classes	Avg. Agents
KITTI	2012	Real	No	15k	15k	200k	8	1
nuScenes	2019	Real	No	1.4M	400k	1.4M	23	1
DAIR-V2X	2021	Real	V2V&I	39k	39k	464k	10	2
V2X-Sim	2021	Simulated	V2V&I	0	10k	26.6k	2	2
OPV2V	2022	Simulated	V2V	44k	11k	230k	1	3
DOLPHINS	2022	Simulated	V2V&I	42k	42k	293k	3	3
V2V4Real	2023	Real	V2V	40k	20k	240k	5	2
WHALES (Ours)	2024	Simulated	V2V&I	70k	17k	2.01M	3	8.4

Agent Types

Location	Category	Sensors	Planning & Control	Tasks	Spawning
On-road	Uncontrolled CAV	LiDAR ×1 + Camera ×4	CARLA autopilot	Perception	Random / deterministic
On-road	Controlled CAV	LiDAR ×1 + Camera ×4	RL policy	Perception & planning	Random / deterministic
Roadside	RSU	LiDAR ×1 + Camera ×4	RL policy	Perception & planning	Static
Anywhere	Obstacle agent	–	CARLA autopilot	–	Random

Getting Started

Installation

1. Clone the repository:

   git clone https://github.com/chensiweiTHU/WHALES.git

2. Create and activate a Conda environment:

   conda create -n whales python=3.10 -y
   conda activate whales

3. Install WHALES:

   pip install -e .

4. Install mmdetection3d==0.17.1 following the official guide.
5. (Optional) Install OpenCOOD for additional cooperative baselines.

Data Preparation

1. Download the full dataset from Google Drive: Download Whales.
2. Place extracted files under ./data/whales/.
3. Preprocess:

   python tools/create_data.py whales --root-path ./data/whales/ --out-dir ./data/whales/ --extra-tag whales

   This emits, under ./data/whales/:
   • whales_infos_{train,val}.pkl — LiDAR info PKLs for WhalesDataset.
   • whales_infos_{train,val}_mono3d.coco.json — per-camera mono3D COCO files for WhalesMonoDataset (cam-only training).
   • whales_dbinfos_train.pkl + whales_gt_database/ — GT-sampling database used by LiDAR configs' augmentation step.

Training & Evaluation

Configs are organised as:

• ./configs/_base_/ — shared dataset, model, and schedule bases.
• ./configs/standalone/ — single-agent baselines (PointPillars, SECOND, CenterPoint, FCOS3D, VoxelNeXt, etc. on LiDAR and monocular 3D).
• ./configs/cooperative/ — V2X cooperative-perception recipes (PointPillars, VoxelNeXt, BEVFusion, FCooper, V2VNet, V2X-ViT, OPV2V, FFNet, plus the scheduling studies).

Pick any leaf config under those trees and run:

• Training

  bash tools/dist_train.sh <config>.py <gpu_num>

• Testing

  bash tools/dist_test.sh <config>.py <model>.pth <gpu_num> --eval bbox

Both WhalesDataset (LiDAR) and WhalesMonoDataset (monocular 3D) are registered; the COCO JSONs emitted by the preprocessing step drive the mono3D path, the info PKLs drive the LiDAR path. Metrics: mAP and NDS.

Visualization

tools/misc/visualize_whales.py can render from all three data representations (raw frame_info.json, info PKL, mono3D COCO):

# Raw CARLA frame_info.json: reconstruct ego-frame boxes + overlay on the 4 cameras + BEV.
python tools/misc/visualize_whales.py frame_info \
    --path data/whales/<scene>/<frame>/frame_info.json --agent vehicle0

# Single entry from the info PKL (one agent-frame).
python tools/misc/visualize_whales.py pkl \
    --path data/whales/whales_infos_val.pkl --token <scene>_<frame>_<agent>

# Batched renders: 2x2 camera grid alongside the BEV, one frame per scene.
python tools/misc/visualize_whales.py pkl_grid \
    --pkls data/whales/whales_infos_{train,val}.pkl \
    --num-per-pkl 20 --one-per-scene --out whales_vis/

# Mono3D COCO renders with 2D bbox + 3D wireframe per annotation.
python tools/misc/visualize_whales.py coco \
    --path data/whales/whales_infos_val_mono3d.coco.json --image-id <image_id>
python tools/misc/visualize_whales.py coco_batch \
    --path data/whales/whales_infos_val_mono3d.coco.json \
    --num-tokens 20 --one-per-scene --out whales_vis_coco/

Scheduling Algorithms

Agent scheduling pipelines live in ./mmdet3d_plugin/datasets/pipelines/cooperative_perception.py. CAHS prioritizes collaborators by historical coverage and predicted gains.

Experimental Results

All numbers below are reported as 50m / 100m, the two evaluation ranges used by the WHALES protocol (per-class radial distance from ego).

Stand-alone 3D Object Detection

Method	AP_Veh ↑	AP_Ped ↑	AP_Cyc ↑	mAP ↑
PointPillars	67.1 / 41.5	38.0 / 6.3	37.3 / 11.6	47.5 / 19.8
SECOND	58.5 / 38.8	27.1 / 12.1	24.1 / 12.9	36.6 / 21.2
RegNet	66.9 / 42.3	38.7 / 8.4	32.9 / 11.7	46.2 / 20.8
VoxelNeXt	64.7 / 42.3	52.2 / 27.4	35.9 / 9.0	50.9 / 26.2

Cooperative 3D Object Detection

Method	AP_Veh ↑	AP_Ped ↑	AP_Cyc ↑	mAP ↑
No Fusion	67.1 / 41.5	38.0 / 6.3	37.3 / 11.6	47.5 / 19.8
F-Cooper	75.4 / 52.8	50.1 / 9.1	44.7 / 20.4	56.8 / 27.4
Raw-level Fusion	71.3 / 48.9	38.1 / 8.5	40.7 / 16.3	50.0 / 24.6
VoxelNeXt	71.5 / 50.6	60.1 / 35.4	47.6 / 21.9	59.7 / 35.9

Scheduling Studies — Single-Agent Policies

mAP at 50m / 100m. Base detector: VoxelNeXt (LiDAR cooperative). Rows = inference-time policy, columns = training-time policy.

Inference \ Training	No Fusion	Closest First	Single Random	Multiple Random	Full Communication
No Fusion (Baseline)	50.9 / 26.2	50.9 / 23.3	51.3 / 25.3	50.3 / 22.9	45.6 / 18.8
Closest First	39.9 / 20.3	58.4 / 30.2	58.3 / 32.6	57.3 / 30.5	55.4 / 10.8
Single Random	43.3 / 22.8	57.9 / 31.0	58.4 / 33.3	57.7 / 31.4	55.0 / 14.6
MASS	55.5 / 11.0	58.8 / 33.7	58.9 / 34.0	57.3 / 32.3	54.1 / 27.4
CAHS (Proposed)	56.1 / 29.6	62.5 / 31.7	62.7 / 35.9	58.3 / 32.6	59.9 / 31.0

Scheduling Studies — Multi-Agent Policies

mAP at 50m / 100m. Base detector: VoxelNeXt (LiDAR cooperative). Same axes as above.

Inference \ Training	No Fusion	Closest First	Single Random	Multiple Random	Full Communication
Multiple Random	34.5 / 16.9	60.7 / 35.1	61.2 / 37.1	61.4 / 36.4	58.8 / 12.9
Full Communication	29.1 / 10.5	63.7 / 38.4	63.7 / 39.1	64.0 / 41.1	65.1 / 39.2
MASS	54.6 / 13.4	64.9 / 39.7	65.0 / 40.5	63.7 / 40.4	63.5 / 36.4
CAHS (Proposed)	53.7 / 14.2	65.3 / 40.1	65.1 / 42.0	63.9 / 40.6	65.2 / 39.2

Roadmap

• Publish dataset and checkpoints on HuggingFace.

Citation

@INPROCEEDINGS{11247472,
  author    = {Wang, Yinsong Richard and Chen, Siwei and Song, Ziyi and Zhou, Sheng},
  title     = {{WHALES: A Multi-Agent Scheduling Dataset for Enhanced Cooperation in Autonomous Driving}},
  booktitle = {2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2025},
  pages     = {20487-20493},
  keywords  = {Wireless communication; Three-dimensional displays; Scalability; Whales; Benchmark testing; Metadata; Scheduling; Vehicle dynamics; Vehicle-to-everything; Autonomous vehicles},
  doi       = {10.1109/IROS60139.2025.11247472}
}