Robot Self-Collision Checking Dataset with Input Positional Encoding for Machine Learning Models

This dataset accompanies the article "Improving Machine Learning-Based Robot Self-Collision Checking with Input Positional Encoding" published in Foundations of Computing and Decision Sciences (DOI: 10.2478/fcds-2025-0015) and contains two benchmark datasets used for binary classification experiments: (1) a synthetic 2D dataset and (2) a Robot40 self-collision dataset.

The Synthetic 2D dataset is an artificially generated benchmark designed to evaluate classification methods in a controlled, low-dimensional setting. It contains 100,000 samples. Each sample consists of:
- Two continuous input variables (x1, x2), both defined in the range [0, 1],
- One binary class label (0/1).
The dataset provides a simple yet nontrivial decision boundary suitable for testing classification algorithms, uncertainty modeling, and visualization techniques. Due to its normalized input space and binary output, it is particularly useful for methodological comparisons and reproducible experiments.

The Robot40 dataset contains kinematic configurations of a 6-DOF robotic manipulator (Universal Robots UR5) mounted on a robotic platform (Robot40). Each data sample includes:
- Six joint angle values (q1–q6), each expressed in radians and defined in the range [−π, π],
- One binary label indicating the presence (1) or absence (0) of self-collision.
Self-collision labels were generated using the Flexible Collision Library (FCL), a widely used library for proximity queries and collision detection in robotics and motion planning.
The Robot40 dataset is provided in multiple size variants to support scalability studies and performance benchmarking. The available versions contain:
- 100,000 samples
- 200,000 samples
- 500,000 samples
- 1,000,000 samples
- 2,000,000 samples
- 5,000,000 samples

Each variant follows the same structure and labeling procedure. The dataset represents a supervised learning problem where the task is to predict self-collision based solely on joint configuration. Since all joint values are sampled within their full rotational range, the dataset covers a broad portion of the manipulator’s configuration space.
In our experiments, the raw input data from both datasets were also optionally transformed using a positional encoding function. This allowed us to analyze how frequency-based feature mappings influence classification performance compared to models trained directly on the original input variables.

Both datasets are provided in a structured format suitable for direct use in common data analysis and machine learning frameworks.