Seite - 80 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Toward Safe Perception in Human-Robot Interaction Inka Brijacak, Saeed Yahyanejad, Bernhard Reiterer and Michael Hofbaur1 Abstract—Perception is a major component ofa system when it comes to the concept of safety in human-robot interaction. Althoughdesigningamechanicallysaferobotmayreduce lotsof potential hazards, it is still beneficiary or even required to have detailed knowledge of the current status of the robot, human, and other environmental entities. We refer to this knowledge as perceptionalawareness,orsimplyperception, that subsumes: (i) whatoursystemperceives fromrobot stateand its environment, (ii) what our system perceives from human state, and (iii) what a human perceives from the robot state. In this paper we provide requirements for a holistic architecture to construct safe perception using multiple heterogeneous and independent sensors and processing units in any environment that includes both robots and humans. We also illustrate our concepts on the basis of particular instances of this scheme realized in the robotic lab. I. INTRODUCTION Nowadays, robots are being used widely in different fields due to their precision, accuracy, reliability, and easy deployment. In many initial applications of robots, they are functioning separated from humans in isolated areas. With advances of technology and the necessity for coexistence of robots and humans (e.g., medical application, service robots, collaborative production lines), the new era of human-robot interaction (HRI) has emerged. HRI studies and describes the types and characteristics of the possible interactions that can exist between a robot and a human. When a human is working in a close distance with robots, the safetyof thehuman becomesan important issue. Initially, safety requirements for many industrial robotic applications were achieved just by a physical separation of humans from any robot (e.g., using barriers or fences). This simple and effective way to impose safety, however, prevents direct interaction between humans and robots to work collabo- ratively. The relevant international standard for safety in industrial robots [10], [11], which specifies accepted means to impose safety, however, allows also human-robot collab- oration in four clearly defined scenarios. The new technical specification ISO TS 15066 “Robots and robotic devices – Collaborative robots” [12] provides even more details on these operational settings and specifies comprehensive force, pressure, and speed limits for unintended human-robot interactions (collisions). Risk reduction during human-robot interaction has three main approaches: (i) redesigning the system and the task realization, (ii) using functional or physical safeguards, and (iii) raising the awareness of the operator/user, either using 1 All authors are with JOANNEUM RESEARCH ROBOTICS - Insti- tute for Robotics and Mechatronics - Cognitive Robotics Group, Austria <firstname.lastname>@joanneum.at active warnings during operation and/or by specific training. Taken into account the industrial experience, redesigning the system is the most effective risk reduction strategy and should always be applied first. However, when operating adaptively in less structured environments, redesign alone is often insufficient, and additional functional safety measures are mandatory [13]. It is possible to combine these three approaches to achieve higher levels of safety. In spite of that, no matter how accurateasystemisdesigned, thecontinuousmonitoring (the second approach mentioned above) is an important factor for a safe system. To be able to understand the status of the environment or a system, the concept of perception plays an important role. Similar to human perception, the concept of the perception for a system can be twofold: • External: What a system sees, perceives, or understands from the environment, i.e., what types of object are around me? What are their positions, speed, shape, size? What are the states of other systems around me? • Internal: What a system sees, perceives, or understands about itself, i.e., where should I be? Where am I? What is my current state? For both of these perception types, we need dedicated sensors to obtain relevant data upon for perception. This demanding task requires to deal with the following issues: • sensory data acquisition and storing the data • data mining, enhancement, and filtering • sensor fusion • time synchronization • dependable, safety-enabled operation. The complexity highly increases with the larger number of heterogeneous sensors such as, safety-enabled laser scanners (LIDARs), RGB cameras, thermal cameras, time-of-flight (ToF) cameras, haptic sensors, proximity sensors, ultrasonic sensors, robot internal sensors (e.g., torque sensors), pressure sensors, etc. Redundancy achieved by using diverse sensor types highly improves the reliability of the overall perception unit. Dealing with diverse sensors requires one to carefully consider the different interfaces, data types, sampling rates and, of course, a potentially large amount of data. In order to deploy such an inclusive perception scheme in real-world robotic systems, however, it is also important to consider the requirements set by the relevant standards that include the entire life cycle of the system starting with the devel- opment process itself, hard- and software-requirements and functional issues for all forms of the system’s operation. This goes far beyond the requirements necessary to realize a laboratory demonstrator. As a consequence, it is helpful to 80