Transcript Slides
Robot Programming from Demonstration, Feedback and Transfer Yoan Mollard, Thibaut Munzer, Andrea Baisero, Marc Toussaint, Manuel Lopes Robot Programming • This paper presents a novel approach for robot instruction for assembly tasks (i.e. teaching a robot how to assemble an object): the usage of a GUI allows users to refine a routine learned through physical demonstration. • A further advantage afforded by this approach is the possibility of knowledge transfer, where similar tasks can be learned through the same refinement process General Goals of Robot Programming • Easy to program: a simple demonstration of a task allows, in most cases, to program it • Easy to instruct: corrections of what has been learned are easy to provide • Shared task awareness: a joint comprehension of the task being executed, the world state and the role of each partner, is shared • Shared initiative: the robot does not require a step-by-step instruction, the system is able to start its own behavior and wait when it is uncertain about the task, allowing the user to correct or change the behavior at any point in time • Hierarchical learning: all parts of the task are easy to program Fields of Interest • Self-improvement and scalability • Training the user to provide better demonstrations • Correcting the task by the means of a visual programming language, allowing the system to request specific demonstrations or clarifications. Demonstration, Feedback & Transfer • Demonstration – A human demonstrator provides visual input which is processed into a series of constraints • Key term: Constraint – Consider tasks as sequences of kinematic constraints, a single constraint is modeled as the transformation between two object frames • Key term: Symbol – Different works in robot demonstration learn symbols differently – High level approach to symbols label key features of a state, or the meaning of instruction signals – Low level approach to symbols label individual actions that are individually taught to a robot – For purpose of this paper, symbols represent individual actions applied to a single object Demonstration, Feedback & Transfer • Feedback – Refine the plan generated from demonstration 1) Constraint representation and correction 2) Assembly plan representation and correction • Problem: how to effectively allow user to correct said plan? – Use a GUI Demonstration, Feedback & Transfer Demonstration, Feedback & Transfer • Transfer – The process of refinement extends from simply the task at hand – Similar tasks should be easily programmable – Correct the differences using the same GUI General Workflow Does it actually work? • Experiment: • Terms – Bootstrap: Rely on a previous demonstration to learn the task and refine it – Cold start: Start directly from task refining with no previous demonstration Demonstration, Feedback & Knowledge Transfer • Demonstration & Feedback – https://vimeo.com/120726868 • Knowledge Transfer – https://vimeo.com/109165300 Does it actually work? Brief explanation of results: Perfect system would create 5 symbols associated with 5 actions after Task Learning (Demonstration) False negatives during the constraint detection phase causes some missing symbols and leads to a missing number of actions, seen above. Does it actually work? Possible Improvements • Consider actions that do not necessarily join two items – E.g. Setting the seat upright before attaching the back, moving objects closer to the robot • Ability to observe multiple humans cooperating and formulate concurrent, sequential or dependent actions • Considering objects by shape rather than attaching all objects into a single frame could lead to more robust constraints (non-rigid constraints)