When to stop for safe manipulation in unstructured environments?

Since robots are getting used in domestic environments, relevant safety measures become crucial for safely interacting with the environment. However, these measures should not interrupt the robot execution persistently or reduce the capabilities of task execution. This can be achieved by a proper decision mechanism that is selective to what to react considering the relevancy to the current task. In this study, we propose an execution monitoring architecture that addresses these issues. This architecture provides the reactions depending on the task in execution, the faced failures and the properties of interactions between the robot and the environment. In particular, we are interested in when the robot should choose to stop the execution as a reaction. We model the problem as a classification problem, and use a neural network based approach. We evaluate the accuracy of the reactions of our humanoid robot in real-world tabletop manipulation scenarios. The results indicate that our architecture can make a decision with 98% accuracy.