Distance-Aware Inverse Reinforcement Learning Dataset for Robotic Grasping
Description
This dataset contains state-action observations collected from a low-cost robotic manipulator during expert-guided demonstrations for robotic grasping. Each observation records four joint measurements, the scanning-servo angle of the time-of-flight (ToF) distance sensor, the measured distance between the robot and the target object, and the corresponding action label. The dataset is intended to support research on inverse reinforcement learning, learning from demonstration, robotic action selection, and resource-constrained autonomous robotic grasping.
Files
Steps to reproduce
1. Download Grasp_dataset.csv from the dataset record. 2. Load the CSV file using a Python, MATLAB, or similar data-analysis environment. 3. Separate the observations into the training set (219 trajectories, 30,238 observations) and testing set (55 trajectories, 7,522 observations). 4. Use the joint measurements, object distance, and action labels to construct the robotic state and feature representations described in the associated study. 5. Train the Standard MaxEnt IRL model without the distance feature and the proposed Distance-Aware MaxEnt IRL model with the distance feature. 6. Evaluate action-selection performance on the test trajectories using accuracy and class-wise/Macro F1 scores.