Synthetic Object Dataset for Affordance-Based Manipulation

This dataset contains synthetic images of everyday objects organized according to morphology-driven manipulation affordances. It was designed to support the evaluation of vision-based methods for semiautonomous upper-limb prosthetic control, robotic manipulation, and affordance-based object understanding.

The curated dataset includes 180 synthetic PNG images, divided into 9 affordance classes, with 20 images per class. Each image depicts a single everyday object in a tabletop scene. The images introduce variability in object instance, material, color, surface, context, camera viewpoint, framing, object position, orientation, and scale.

Unlike conventional object-recognition datasets, labels are not assigned only according to object identity. Instead, each object is categorized according to the manipulation strategy suggested by its shape, size, and grasp-relevant visual properties. For example, bottles, thermos flasks, and cylindrical containers are grouped together because they suggest a similar power grasp, while coins, buttons, and small screws are grouped as small objects because they are typically manipulated through precision pinch.

Each class folder contains 20 images. Folder names correspond to the affordance label assigned to the images, while filenames identify the object instance represented in each image.

The dataset can be used for affordance-based image classification, evaluation of open-vocabulary or zero-shot vision models, manipulation-oriented semantic recognition, comparison between object-identity labels and grasp-affordance labels, and preliminary testing of perception modules for prosthetic or robotic manipulation pipelines.

All images are synthetic. This enables controlled evaluation under systematic variations of object appearance and scene conditions. However, the dataset does not fully capture real-world variability, clutter, occlusions, complex lighting conditions, or sensor noise. For this reason, models evaluated on this dataset should also be validated on real-world images before being used in practical prosthetic or robotic systems.

This release includes the curated image dataset organized in class folders. Auxiliary generation materials, prompt-based outputs, and the image generation pipeline are not intended as part of the main benchmark dataset.