Robin, just under a year ago, a robot in Russia broke a child's finger in a game of chess. The robot arm accidentally grabbed not the chess piece but one of the boy's fingers. Could something like this also happen with your robot arms?
With our work we make sure incidents like that cannot occur. Our robot arms are tactile, unlike the robot arm that moved the chess pieces. They sense contact with an obstacle and react to it accordingly, e.g., by stopping. In our research, we are concerned with two main questions. On the one hand, we investigate the occurrence of human injury, so to say “how high may contact forces be without causing injuries?”. On the other hand, we are concerned with finding out how precisely the robot can measure, adapt to, and apply forces: “Can my robot control and measure forces well enough to comply to safety force thresholds and ensure reliable task execution?”. For a comparison of the systems, we use test procedures and setups with which we can test various robot arms from different manufacturers. These benchmarks can then be used, for example, in our SafeRoBAY human-robot collaboration (HRC) safety research project to enable biomechanically safe and efficient HRC.
New metrics: How sentient is a robot?
What does that look like in practice?
Currently, for example, we have attached a sphere to this robot arm from Kuka that presses on a force plate with a predefined force. A force controller tells the robot the motor torques needed to achieve a desired force, and torque sensors built into each of the seven joints are designed to help it estimate the actual force being generated. My question is: If we ask the system to apply a predefined force, how accurately will it reach this force? There are many factors that influence how exactly the actual force corresponds to the desired one. The sensing technology that is applied for contact force estimation for example. With systems using contact force estimation via motor current readings we see a higher deviation from the desired force and a larger effect of the robot run time on the measurement than for robots using torque sensors, like this Kuka robot. As the main part of my doctoral thesis, I develop a standardized set of basic metrics, benchmarking procedures, and test setups required to describe robot tactility and to transfer them into a model that allows me to estimate the performance of force control of robot arms. In order to then find out whether injuries can still occur at all in the HRC with the robot systems that show high tactility in our benchmark tests, we use both standardized test devices and biomechanical injury tests as part of the SafeRoBAY project. Altogether, this results in a basic set of metrics for robot tactility and a tool for predicting the necessary safety measures for HRC - our big formula for safe force-controlled robots.
What are the differences between the robot arms?
The handling, mechanics and control methods are very different. You can already feel this when you guide the robot by hand. The robot arms from Kuka and Franka, for example, which have the higher tactility than most other systems here, feel very smooth-running, others less so. In addition, we have found in studies that the risk of injuries in HRC with more tactile systems, for example from the Franka Emika robots, is very low. We currently have a set of more than 20 metrics, with respective measurement setups and procedures that we can use to quantify the differences in tactility and safety of the robotic systems. By now, we tested 11 different robot systems (publication upcoming) and in the meantime, we have become quite experienced in this field, so that the first industrial customers are already having their robot arms tested by us.
Selected publications by Robin Kirschner
Towards a Reference Framework for Tactile Robot Performance and Safety Benchmarking, 2021
Manual Maneuverability: Metrics for Analysing and Benchmarking Kinesthetic Robot Guidance, 2022