Safety First: a Dataset of Harmful Task Plans for Robots

Zainab Altaweel, Mohaiminul Al Nahian, Isaac Lehrer, Adnan Siraj Rakin, Shiqi Zhang

Keywords: Safety in Robotics, Robot task planning, Harmful plans dataset

TL;DR: Introducing a harmful task plans dataset to expose vulnerabilities in robot task planning and enable the development of safety mechanisms against harmful behaviors

Abstract: Ensuring safety in robot task planning is a critical challenge, especially as robots become increasingly integrated into our daily life. In this work, we introduce a dataset of harmful task plans for robots, it is designed to expose vulnerabilities in planning frameworks and highlight overlooked safety concerns. The dataset is generated by injecting base task plans with harmful behaviors, and it is categorized into three levels of detection difficulty. The goal of this dataset is to lay the groundwork for developing robust safety mechanisms in robots. This contribution aims to inspire future research in risk detection and the creation of safer, more reliable task planning systems.

Submission Number: 16