Seite - 157 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

4.3. Findobject This functionrequiresnavigation toasetofpredefinedsearchingpositions,usually located infrontof tables, chestsofdrawersandotherhorizontal surfacesat intermediateheight intervals. Thesearching positionswereplacedaround60-80cmawayfromthesurfacesborder, trying tocovera largeareaof thehorizontalplanewith theheadsensor. For this function, thepath toall thesearchingpositionswas requested by setting a navigation task and muting the navigation until a path is received or a given timeout is reached. This way, it was possible to obtain the path from MIRA, which does not allow arbitrary path planning. If all paths are received, the searching positions are checked in increased lengthorder, otherwisegoals towhicha path isnotprovided arevisited first. 4.4. Call robot At first, a fixed predefined place was associated to each call button id, so that the robot would come to that place when the call button was pressed. This simple method was not flexible enough, hence a novel approachwasdeveloped for this function. It was convenient for the users to call the robot while sitting on their favourite chairs, armchairs, sofas,beds,etc. However, since theprototypeplatformwasnotable torotateon thespotand theback side of the robot could collide with the furniture when turning around to drive away, the predefined positions could not be very close to the actual sitting place. It is also important to take into account that there can be different localization errors, plus the allowed error to decide whether the goal has been reached, so the real position is not always exactly the same (the uncertainty accumulates). A compromisewasoftenneededsothat theusercouldreachthetouchscreenwhileallowingtherobot to detect theobstacleanddriveawaysafely. Furthermore, amoresignificant limitationwas the fact that chairs and armchairs usually can move and the sitting position of the user along a sofa or a bed will most likely vary from time to time. So the desired final position should not be fixed in an improved method. The new approach we present incorporates user detection and interaction, remembered obstacles and discrete motion commands for coming closer to the user with better, adapted positioning. Discrete motioncommands towards thedetecteduserwerechosenoveraplannedpath for threemain reasons: 1) the movement is more direct and predictable; 2) with our sensor setup, obstacle avoidance while followingthepathwouldrequire theheadto lookdownwhereasuserdetectionandanicer interaction require the head to look straight forward; 3) existing navigation frameworks, to our knowledge, are not directly suitable for planning a path to a goal out of the fov or blocked and inside an occupied area,sodefiningapropergoalposeinfreespacewouldbechallenging,dependingontheenvironment characteristicsandrequiringhigher levelknowledgeabout thespecific furniture (orientation, feasible approachdirectionsandsoforth). Discretemotioncommandsbasedsolelyon theuserdetectionwith noobstacleavoidancewhatsoever, on theotherhand, canbe risky. Our solutionworksas follows. Predefinedpositionsaredefinedfurtherawayfromthesittingzoneso that the user’s skeleton should be inside orclose to the headsensor fov. Themaximum distance limit is given by the defined range of the virtual scan for obstacle avoidance or by the maximum distance foradiscretemotioncommand. Whenapredefinedposition is reached, thenavigationparametersare changedsothatoccupiedareas in thecurrent localmaparenotdegradedoroverwrittenbynewsensor data(usernavigationmode). Thentheheadcanmoveupforuserdetection[15],andray-tracingonthe local map is performed. The local map’s origin is located at the odometry reference system and the 157