LATEX TikZposter Workcell Auto-Label Dataset A Multi-Camera, Multi-Object ROS2 Dataset for Robot Manipulation and Vision Benchmarking Kevin Ortega Dept. Electrical Engineering – GSEEA Research Group Universidad Tecnol´ogica de Pereira Pereira, Colombia Email: kevin.ortega@utp.edu.co Batch 1 Release, April 2026 Workcell Auto-Label Dataset A Multi-Camera, Multi-Object ROS2 Dataset for Robot Manipulation and Vision Benchmarking Kevin Ortega Dept. Electrical Engineering – GSEEA Research Group Universidad Tecnol´ogica de Pereira Pereira, Colombia Email: kevin.ortega@utp.edu.co Batch 1 Release, April 2026 Abstract We present Workcell Auto-Label Batch 1, an open dataset for object detection in industrial robot-manipulation environments. The dataset contains 47,650 annotated frames captured simultaneously from three calibrated cameras mounted around a simulated ROS2 workcell, alongside synchronized UR5 robot joint states. All annotations are generated via an automated labeling pipeline and provided in YOLO format, covering four visually distinct object classes. The dataset is designed to benchmark multi-view detection, domain adaptation, and sim-to-real transfer in manipulation tasks. Quick Stats: 47,650 frames – 20,613 images – 3 cameras – 4 object classes – 138,020 annotations. Introduction Robotic manipulation research requires large, annotated perception datasets that couple visual observations with robot state. Hand-labeling such datasets is expensive and error-prone.

• Built in a ROS2 / Gazebo workcell simulation.

• Three synchronized cameras provide different

viewpoints of the same scene.

• Two UR5 robot arms with joint-state logging at

every frame.

• YOLO-format labels ready for direct training with

Ultralytics YOLOv8/v11.

• Supports detection, tracking, and multi-view fusion

research.

Auto-Labeling Pipeline The pipeline exploits the known simulation state to derive exact pixel-level bounding boxes without any human annotation.

ROS2

Workcell Simulation Multi-Camera Image Capture (3 cameras) Robot Joint State Logging (UR5) Auto-Label Pipeline (YOLO format)

YOLO

Dataset Batch 1 Auto-Labeling Pipeline Overview Fig. 1: Simulation pipeline.

Dataset Structure The dataset is organized into Train/Val/Test splits (80/10/10%).

train

80.1%

val

10.0%

test

9.9%

Dataset Split Distribution (Total: 47,650 frames)

0 objects

(background)

20 objects

(full scene) Objects per Frame

0

1000

2000

3000

4000

5000

6000

7000

Frame Count Object Count Distribution (Frames with Objects) Fig. 2: Left: Train/Val/Test split. Split Frames Images w/ Objects Background Train

38,162

16,581

6,874

31,288

Val

4,758

2,055

859

3,899

Test

4,730

1,977

862

3,868

Total 47,650 20,613

8,595

39,055

Table 1: Dataset splits summary.

Sample Annotated Frames The same scene is captured simultaneously by three cameras. Bounding boxes are colored by class: red = cube red, green = cube green, blue = cube blue, gray = sphere gray.

Camera 1 (Front) Camera 2 (Side) Camera 3 (Top) Multi-Camera Auto-Labeled Frame Workcell Scene Fig. 3: Multi-camera synchronized capture of the same scene from three viewpoints.

CAM1

frame04077...

CAM3

frame04078...

CAM2

frame04078...

CAM1

frame04078...

CAM3

frame04079...

CAM2

frame04079...

Sample Auto-Labeled Frames (Train Split) cube\_red cube\_green cube\_blue sphere\_gray Fig. 4: Six representative labeled frames from the training split across all three camera viewpoints, illustrating scene variability. Object Classes and Annotations cube\_red cube\_green cube\_blue sphere\_gray

0

10000

20000

30000

40000

Annotation Count

13,802

41,406

41,406

41,406

Object Class Distribution Across All Splits Fig. 5: Class distribution of annotations across the dataset. Bounding Box Statistics

2.6

2.7

2.8

2.9

3.0

3.1

3.2

3.3

Bounding Box Width (%)

3.4

3.6

3.8

4.0

4.2

4.4

Bounding Box Height (%) BBox Size Distribution by Class cube_red cube_green cube_blue sphere_gray

0.09

0.10

0.11

0.12

0.13

0.14

BBox Area (% of image)

0

5000

10000

15000

20000

25000

Count Bounding Box Area Histogram Fig. 6: Left: width vs. height scatter (% of image dimensions) by class. Right: bounding-box area histogram. Key geometric properties:

• Width range: 1.8%–5.5% of image width.

Synchronized Robot State Each frame record in annotations.csv includes 6- DOF joint angles (radians) for both robot arms: joints_ur1: [θ1, θ2, θ3, θ4, θ5, θ6] joints_ur2: [θ1, θ2, θ3, θ4, θ5, θ6] This makes the dataset suitable for:

• Kinematic-aware detection: arm pose as context

feature.

• Occlusion prediction: arm linkage models predict

self-occlusion.

• Grasp outcome studies: correlating visual state

with arm configuration.

Use Cases & Applications

• Object detection baselines: Plug-and-play with

YOLOv8/v11, RT-DETR, etc.

• Multi-view fusion: Three synchronized views en-

able stereo/trinocular research.

• Sim-to-real transfer: Evaluate domain adaptation

with real counterpart data.

• Hard negative mining: 39K background frames for

false-positive reduction.

• Robot learning:

Joint state + image pairs for vision-based control policies.

Conclusions Workcell Auto-Label Batch 1 provides a large-scale, automatically annotated multi-camera dataset for industrial robot-manipulation scenes. Key contributions:

1. An end-to-end auto-labeling pipeline requiring

zero human annotation effort.

2. 20,613 images across 3 viewpoints with 138,020

bounding boxes.

3. Synchronized robot joint-state metadata enabling

state-conditioned models.

4. Immediate

compatibility with modern

YOLO-

family detectors.

Dataset Access Format: YOLO (txt labels + JPEG images) Config: dataset.yaml (nc=4, train/val/test splits) Contact: kevin.ortega@utp.edu.co