Human-Robot Collision Avoidance Scheme for Industrial Settings Based on Injury Classification Mustafa Mohammed University of Illinois at Chicago Chicago, IL, United States mmoham70@uic.edu Heejin Jeong University of Illinois at Chicago Chicago, IL, United States heejinj@uic.edu Jae Yeol Lee Chonnam National University Gwangju, Republic of Korea jaeyeol@chonnam.ac.kr ABSTRACT The objective of this paper is to develop a real-time, depth- sensing surveillance method to be used in factories that require human operators to complete tasks alongside collaborative robots. Traditionally, collision detection and analysis have been achieved with extra sensors that are attached to the robot to detect torque or current. In this study, a novel method using 3D object detection and raw 3D point cloud data is proposed to ensure safety by deriving the change in distance between humans and robots from depth maps. By not having to deal with any potential delay associated with extra sensor-based data, both the likelihood and severity of collaborative robot-induced injuries are expected to decrease. CCS CONCEPTS • Human-centered computing → Human-computer interaction (HCI) → Interaction paradigms → Collaborative interaction; KEYWORDS Collaborative robot; point cloud; computer vision; collision detection; safety; injury prevention ACM Reference format: Mustafa Mohammed, Heejin Jeong and Jae Yeol Lee. 2021. Human-Robot Collision Avoidance Scheme for Industrial Settings Based on Injury Classification. In Companion of 2021 ACM/IEEE International Conference on Human-Robot Interaction (HRI’21 Companion), March 8-11, 2021, Boulder, CO, USA. ACM, New York, NY, USA, 3 pages. https://doi.org/10.1145/3434074.3447232 1 Introduction The emergence of collaborative robots into human workspaces has induced a complete reevaluation of the way service tasks are handled. The ingenuity of humans is freed from its limitations by robots that operate with precision and speed. Combining the two entities into industrial settings has allowed for not only a surge in efficiency but also a drop in the occurrence of adverse effects from physical labor. However, to take full advantage of the initiative, it is critical that any added harm must not be inflicted upon factory workers during manufacturing and assembling. To date, an abundance of injury prevention methods relying on marker-based sensors has been tested [1]. These sensors are installed into the robot; thus, the approach can introduce delay into the rate at which the robot is stopped, which can pose an immediate threat to workers [1]. A non-marker-based approach is therefore imperative. There have been numerous proposed algorithms that utilize active image-based technologies (both 2D and 3D) to solve the issue of collaborative robot safety. Saveriano and Lee [2] suggested a real-time path planning algorithm for reactive avoidance of multiple moving obstacles with computer vision. From their simulations, they were able to achieve a significantly shorter and collision-free travel path. In regard to environment mapping, Rusu et al. [3] proposed a voxel-based sensing method that created obstacle maps with the application of point clouds. Himmelsbach et al. [4] presented a safer approach to speed and separation monitoring with 3D Time-of-Flight (TOF) cameras that were directly mounted on a robot arm. Moreover, an efficient method to determine the distance between obstacles in the path of a robot was developed by Flacco et al. [5]. With the use of multiple depth cameras, they generated repulsive vectors that guided the robot controller through a task. Lešo et al. [6] focused on developing 2D safety zones from projected optical barriers that, if crossed, would lead to the stoppage of the robot. Though not as common, Cherubini et al. [7] opted for a multimodal approach, in which both traditional and depth cameras were simultaneously utilized to trigger safety stops in a manufacturing cell. Most importantly, experiments done by Stetco et al. [8] clearly demonstrate the advantage of using 3D techniques over all others. Furthermore, many have examined methods for quantifying injury risk, especially with collaborative robots. Matthias et al. [9] worked on understanding low-level injury risk assessment by matching the area of human-robot contact with injury type. Robla-Gómez et al. [10] reviewed a framework that used impact energy density. In the chance of a collision, an altered trajectory for the robot is generated. Marvel and Norcross [11] tested an algorithm that takes the angle of impact and contextualizes it with the separation distance between the human operator and the robot. In addition to this, Haddadin et al. [12] explored injury Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from Permissions@acm.org. HRI ’21 Companion, March 8-11, 2021, Boulder, CO, USA © 2021 Association for Computing Machinery. ACM ISBN 978-1-4503-8290-8/21/03…$15.00 DOI: https://doi.org/10.1145/3434074.3447232 Late-Breaking Report HRI ’21 Companion, March 8–11, 2021, Boulder, CO, USA 549