Vision-based grasping of unknown objects to improve disabled persons autonomy

摘要

Abstract —This paper presents our contribution to vision basedrobotic assistance for people with disabilities. The rehabilitativerobotic arms currently available on the market are directlycontrolled by adaptive devices, which lead to increasing strainon the user's disability. To reduce the need for user's actions,we propose here several vision-based solutions to automatize thegrasping of unknown objects. Neither appearance data basesnor object models are considered. All the needed informationis computed on line. This paper focuses on the positioning of thecamera and the gripper approach. For each of those two steps,two alternative solutions are provided. All the methods have beentested and validated on robotics cells. Some have already beenintegrated into our mobile robot SAM . I. I NTRODUCTION This work relates to robotic assistance for disable people,where autonomous robotic systems are designed to compen-sate for a human motor disability. We propose solutions forthe grasping of any object within a domestic environment suchas an apartment. Providing a robust, generic and easy-to-usesolution to improve the user's interaction with their personalenvironment would largely increase their autonomy.Contrary to an industrial environment [19], the domesticenvironment is highly unstructured. Thus, the robotic systemneeds exterioceptive sensors to adapt its behavior to the currentsituation. Vision sensors are almost always used: this sensoris quite cheap, the acquired information is very rich, and itcan even be directly used as feedback for the user.A. State of the artBefore starting a grasping procedure, a robotic system rstneeds to extract information on the object from the visualinput. In order to handle any object shape and appearance, itis necessary to make some assumptions on the situations therobot can handle.Some approaches propose to constrain the possible locationsfor the object. For example, [11] assumes that the scene isknown and uses a simple image difference with the knownbackground to localize the object. The project