Contact-State Segmentation Using Particle Filters for Programming by Human Demonstration in Compliant-Motion Tasks

This paper presents a contribution to programming by human demonstration, in the context of compliant-motion task specification for sensor-controlled robot systems that physically interact with the environment. One wants to learn about the geometric parameters of the task and segment the total motion executed by the human into subtasks for the robot, that can each be executed with simple compliant-motion task specifications. The motion of the human demonstration tool is sensed with a 3-D camera, and the interaction with the environment is sensed with a force sensor in the human demonstration tool. Both measurements are uncertain, and do not give direct information about the geometric parameters of the contacting surfaces, or about the contact formations (CFs) encountered during the human demonstration. The paper uses a Bayesian sequential Monte Carlo method (also known as a particle filter) to do the simultaneous estimation of the CF (discrete information) and the geometric parameters (continuous information). The simultaneous CF segmentation and the geometric parameter estimation are helped by the availability of a contact state graph of all possible CFs. The presented approach applies to all compliant-motion tasks involving polyhedral objects with a known geometry, where the uncertain geometric parameters are the poses of the objects. This work improves the state of the art by scaling the contact estimation to all possible contacts, by presenting a prediction step based on the topological information of a contact state graph, and by presenting efficient algorithms that allow the estimation to operate in real time. In real-world experiments, it is shown that the approach is able to discriminate in real time between some 250 different CFs in the graph