Abstract
Subsea Inspection, Maintenance and Repair (IMR) interactions on underwater Oil & Gas infrastructure can have severe consequences in case of failure. Currently, these interactions are mainly carried out using Remotely Operated Vehicles (ROVs) with attached robotic arms where operators assess the situation and make decisions. To allow for increased autonomy in operations on routine objects (valves, wires, hoses, tools), the ROV has to detect the objects and their pose before manipulation tasks can be performed. These tasks typically involve risks, and therefore it is desirable to estimate the probability of operation failure in order to provide decision-support to human operators during the mission. In this paper, we propose a framework using machine learning with a Gaussian Naive Bayes Classifier to estimate the failure-probability of robotic tasks based on the objects' spatial uncertainties. As the uncertainty input-feature we use the 6 DOF standard deviation of the object's pose-estimate.We show how prediction accuracy improves over time and how well the predictions match actual failure rates. We run 1000 simulated pick-and- place operations with different uncertainties and discuss how our method can improve decision-support during operation. We also include a small dataset collected by a 3D camera from real-world objects, test the transferability of simulation results to these data and a pose-estimation algorithm, and examine the impact of data quality.