sayedpedramhaeri/VLA4CoDrive

Vision–Language–Action Dataset for Cooperative Autonomous Driving

VLA4CoDrive is a large-scale cooperative Vision–Language–Action (VLA) dataset designed to support autonomous driving under multi-vehicle cooperation. This work has been accepted to the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2026. This dataset was developed at AI-SENDs Lab, Clemson University, USA.

---

🔍 Overview

We introduce VLA4CoDrive, a cooperative Vision–Language–Action dataset with synchronized multi-vehicle sensing across diverse driving environments, providing multi-view visual streams, contextual text annotations including caption, context, description, and reasoning, and future trajectory actions for training and evaluating VLA driving models.

---

📌 Dataset

VLA4CoDrive is organized into three tightly aligned modalities, Vision, Language, and Action, each captured under the following settings:

🤝 Cooperative Multi-Vehicle Setup Synchronized sensing from multiple vehicles within the same driving episode. 👁️ Multi-View & Multi-Modal Perception RGB front, rear, left, and right cameras, LiDAR, semantic LiDAR, optical flow, GNSS, and IMU. 📝 Structured Vision–Language Grounding Clip-level annotations including Context, Caption, Description, and Reasoning. 🎯 Action & Trajectory Supervision Low-level controls including steer, throttle, and brake, together with 30-step future trajectories.

🌦️ Controlled Diversity 8 CARLA towns × 8 weather conditions with frame-aligned replay. 📏 Large-Scale Dataset ~10M vision samples ~150K language annotations ~1M action records ~300–360 hours of driving data 📦 Standard Annotation Formats COCO, PASCAL VOC, KITTI 2D, and KITTI 3D.

---

👁️ VLA4CoDrive — Vision

Multi-view and multi-modal perception from synchronized cooperating vehicles, capturing complementary visual evidence for cooperative understanding.

---

📝 VLA4CoDrive — Language

Structured, clip-level language grounding generated from synchronized multi-agent scenes, capturing both scene semantics and short-horizon driving intent.

The language annotations include:

• Context
• Caption
• Description
• Reasoning

---

🎯 VLA4CoDrive — Action

Time-aligned action supervision providing low-level controls and future trajectory targets, enabling imitation, forecasting, and language-conditioned planning.

Key	Value
frame
frame_id	002622
timestamp	12481
gearShifter	drive
brake	0.0
brakePressed	false
vEgo	8.3380
vEgoRaw	8.3380
aEgo	2.2104
accelerations_device	[2.2104, -3.0556, 9.8081]
accelerations_calib	[2.2104, -3.0556, 9.8081]
angular_velocities_device	[0.00230, 0.00070, -0.36498]
angular_velocities_calib	[0.00230, 0.00070, -0.36498]
velocities_calib	[8.3296, -0.3735, 0.0000]
positions_ecef	[6378139.50, -48.14, 15.43]
extrinsic_matrix	4×4 matrix (see JSON)
intrinsic_matrix	3×3 matrix (see JSON)
trajectory_count	30
trajectory	[[0.0, 0.0, 0.0], [0.8275010935306759, -0.04821085876190756, 5.1460228860378265e-05], [1.6381458417505876, -0.12501369526440165, 0.00012107845395803452], [2.4244511094026953, -0.2344291940837998, 0.000176658621057868], [3.185244780733905, -0.39312163062116134, 0.00019313814118504524], [3.9298614372630545, -0.612868890162207, 0.00014945981092751026], [4.683324309610262, -0.8823274481627621, 5.035405047237873e-05], [5.454253093046091, -1.1856216876290973, -6.904592737555504e-06], [6.214982731927573, -1.5261434656207276, 4.1389488615095615e-05], [6.948101762385089, -1.8921376276380677, 0.00011508946772664785], [7.644002232229405, -2.2812337434234577, 0.00018184666987508535], [8.311576173004438, -2.6883381010997303, 0.000185737619176507], [8.972671090382748, -3.1091972391105656, 0.00012935639824718237], [9.642792504180107, -3.5487499615343214, 5.081179551780224e-05], [10.324735033181911, -4.005787375082028, 9.19343437999487e-06], [11.00660039244547, -4.473844014640873, 3.288267180323601e-05], [11.676966626025777, -4.941858673255256, 9.204866364598274e-05], [12.332093731260388, -5.4033781985680545, 0.00014080049004405737], [12.976052651337021, -5.859850036152469, 0.00014957424718886614], [13.619010292218759, -6.3176175492788325, 0.00011367793194949627], [14.270971598023293, -6.783213042584521, 5.947111640125513e-05], [14.933883070788191, -7.2576134174966835, 2.8686481527984142e-05], [15.600821812553187, -7.7356100048777305, 4.264828749001026e-05], [16.262800652134572, -8.2105895403186, 8.396152406930923e-05], [16.9149774520535, -8.680089524643837, 0.00012001034338027239], [17.55967605750605, -9.145391107077412, 0.0001282120356336236], [18.203849299595657, -9.610817642782836, 0.00010505679529160261], [18.854042932942548, -10.080827324935955, 6.954197306185961e-05], [19.51155368016565, -10.556230495428062, 4.9018883146345615e-05], [20.171992264576062, -11.033851917102254, 5.714420694857836e-05]]
caption	The ego vehicle is moving straight at a moderate speed following a leading car with acceleration. A nearby traffic light shows green under rainy conditions on a wide road. No pedestrians are present. The driver should remain attentive to the traffic light and be prepared to stop if it changes.

---

🚀 Download Dataset

You can download the dataset using the Hugging Face Hub:

from huggingface_hub import snapshot_download

dataset_path = snapshot_download(
    repo_id="YOUR_USERNAME/VLA4CoDrive",
    repo_type="dataset"
)

print(dataset_path)