The CogMan project (1) develops computational and control models of pick-and-place tasks in the context of everyday manipulation activities in human environments, (2) implements the model into a control system for the kitchen scenario, and (3) empirically analyzes the impact of this control model on the flexibility, robustness, adaptability, and naturality of the robot behavior.
A very impressive aspect of human manipulation of objects in their everyday environments is that they select and parametrize their reaching and grasping movements very skillfully based on the properties of the objects they are to pick up, on the situational circumstances, and on the end positions of the objects. Actions with two collaborators are also adjusted accordingly. The results are very smooth, predictable, and efficient compound movements. Not only do people perform these actions skillfully but they are also capable of learning these skills from little experience and adapt them automatically when needed. Robots that are to serve as robotic assistants and work together with humans need a similar level of manipulation skills and equally powerful means for skill acquisition and adaptation.
Based on experiments analyzing human manipulation tasks from the cognitive psychology point of view, we develop a model of how humans select and parametrize grasps, reaching movements, standing positions, and destinations based on the task context, the object, and the goals of the manipulation activity. Using these information pieces we learn behavior models from the observed data and interpret the findings in the light of cognitive control model.
Using mechanical models of objects, a system can predict what is going to happen to an object when a force is applied to it. A robot can use this knowledge to learn how to exert forces to the object to obtain an specific desired goal. Finding a mapping between this object control and a symbolic representation for it can lead to a language that can be used for robot control.
The novel aspect of the CogMan research approach is that we perform comprehensive experiments of human everyday manipulation activities taking the context into account. Therefore, we use more comprehensive sensor data including estimation of fullbody pose, hand pose, visual attention, various biosignals, model and pose of the manipulated object, force used for lifting, kind of grasp, grasp points, local scene around the object, activity context, etc.
When humans perform even very simple reaching movements, they obey a number of constraints for not colliding, maximizing robustness in the face of uncertainties in perception and actuation. At the same time, the freedoms, that are available in human redundant arms and the task itself are exploited to reduce jerk, execution time and control effort. This optimization takes into account “motor noise” and leads to very high variations in human movements, even under very constrained lab conditions.
One goal of CogMan is to extract these constraints from human motions and represent them in a geometrical manner. Additional constraint- or freedom-'directions' can be taught directly by moving the robot arm, to complete this set of constraints. They are represented using the iTaSC framework, which can combine these contraints into an instantaneous motion controller. It allows us to start from engineered motion controllers and incrementally add constraints and freedoms, and we can check continuously the effectiveness of every added constraint.
This project is partly funded by CoTeSys.
@InProceedings{stulp09compactmodels,
author = {Freek Stulp and Erhan Oztop and Peter Pastor and Michael Beetz and Stefan Schaal},
title = {Compact Models of Motor Primitive Variations for Predictable Reaching and Obstacle Avoidance},
booktitle = {9th IEEE-RAS International Conference on Humanoid Robots},
year = {2009},
bib2html_groups = {Cogman},
bib2html_rescat = {Action},
bib2html_pubtype = {Conference Paper},
}