Industrial Robot Collision Handling in Harsh Environments

• Main Author: Kaldestad, Knut Berg
• Format: Dissertation
• Language: English
• Subjects: eksplosjonsfare; gass; industriroboter; kollisjon; kollisjonshåndtering; olje

The focus in this thesis is on robot collision handling systems, mainly collision detection and collision avoidance for industrial robots operating in harsh environments (e.g. potentially explosive atmospheres found in the oil and gas sector). Collision detection should prevent the robot from colliding and therefore avoid a potential accident. Collision avoidance builds on the concept of collision detection and aims at enabling the robot to find a collision free path circumventing the obstacle and leading to the goal position. The work has been done in collaboration with ABB Process Automation Division with focus on applications in oil and gas. One of the challenges in this work has been to contribute to safer use of industrial robots in potentially explosive environments. One of the main ideas is that a robot should be able to work together with a human as a robotic co-worker on for instance an oil rig. The robot should then perform heavy lifting and precision tasks, while the operator controls the steps of the operation through typically a hand-held interface. In such situations, when the human works alongside with the robot in potentially explosive environments, it is important that the robot has a way of handling collisions. The work in this thesis presents solutions for collision detection in paper A, B and C, thereafter solutions for collision avoidance are presented in paper D and E. Paper A approaches the problem of collision avoidance comparing an expert system and a hidden markov model (HMM) approach. An industrial robot equipped with a laser scanner is used to gather environment data on arbitrary set of points in the work cell. The two methods are used to detect obstacles within the work cell and shows a different set of strengths. The expert system shows an advantage in algorithm performance and the HMM method shows its strength in its ease of learning models of the environment. Paper B builds upon Paper A by incorporating a CAD model of the environment. The CAD model allows for a very fast setup of the expert system where no manual map creation is needed. The HMM can be trained based on the CAD model, which addresses the previous dependency on real sensor data for training purposes. Paper C compares two different world-model representation techniques, namely octrees and point clouds using both a graphics processing unit (GPU) and a central processing unit (CPU). The GPU showed its strength for uncompressed point clouds and high resolution poin