Abstract
Shared control can assist a human tele-operator in performing tasks on a remote robot, but also adds complexity in the user interface to allow the user to select the mode of assistance. This letter presents an expert action recommender framework that learns what actions are helpful to accomplish a task, and generates a minimal set of recommendations for display in the user interface. We address the learning problem in an open world context where the action choice depends on an unknown number of objects, i.e., the output domain of the prediction problem changes dynamically. Using structured prediction, we can simultaneously learn what actions to suggest and what objects those actions should act on. In experiments on three tasks in cluttered table-top environments, this method achieves over 90% accuracy in producing the correct suggestion in the top 5 predictions, and also generalizes well to novel tasks with limited training data.
Highlights
T ELEOPERATION of robotic manipulators has potential to aid a wide variety of applications including searchand-rescue operations, assistive robotics, and remote medical care
We evaluate structured action prediction (SAP) and a combination of featurespace regression with a nearest neighbors search (R+NN) for finding these recommendations
Based on data gathered on a physical robot from cluttered table-top scenes, SAP outperforms R+NN in terms of action selection accuracy across multiple different tasks, achieving top 5 accuracy rates above 90%
Summary
T ELEOPERATION of robotic manipulators has potential to aid a wide variety of applications including searchand-rescue operations, assistive robotics, and remote medical care. Many previous works have used machine learning to infer the operator’s intent and autocomplete tasks [4, 5, 6, 7] Each of these methods performs well in some context, but the variety of contexts in which a robot operates means teleoperators need access to many kinds of actions at different points during operation. The EAR infers the most likely actions an expert user would perform at a given time, taking into account the desired task, the robot’s history of states, and perception information from the environment. It prompts the operator with the k most probable actions, much like a predictive text system on smartphones, so that their decision reduces to only considering which suggestion is best or falling back to the original menu. We find that SAP generalizes well in the few-shot setting on novel tasks, requiring fewer demonstrations to achieve high accuracy once examples from other tasks have been collected
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
