Efficient Models for Grasp Planning With A Object Model Finger - PowerPoint PPT Presentation

Introduction Grasp Frame Efficient Models for Grasp Planning With A Object Model Finger Multi-fingered Hand Workspace Model Grasp Ranking Jean-Philippe Saut, Daniel Sidobre Results Conclusion LAAS-CNRS, FRANCE IROS 2010 Workshop on Grasp Planning and Task Learning by Imitation Taipei, October, 18 th 2010 1/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Introduction Introduction • Goal : computing a dense re-usable grasp set for a Grasp Frame given multi-fingered hand and object. Object Model • Applications : Finger - Manipulation task planning Workspace Model (pick-and-place, dual-hand manipulation) Grasp Ranking Results Conclusion - Interactive grasping (HRI) 2/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Computing a set of grasps vs Introduction computing a unique (“optimal”) Grasp Frame grasp Object Model Finger Workspace Pros Model • The choice of a grasp is very context-dependent and Grasp Ranking this context is susceptible to changes : Results - During pick-and-place tasks, the objects Conclusion are moved by the robot. - In HRI tasks, the human may move the objects. • Re-usable grasp sets allow back-tracking during planning of complex manipulation tasks. Cons • Can be computationally expensive. 3/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Definitions We define a grasp by : Introduction Grasp Frame • A transform between the object and the hand palm Object Model Grasp Frame . Finger Workspace • A set of joint parameters for each finger i : θ i 1 , θ i 2 , . . . , θ i n . Model • A set of contact points ( p 1 , p 2 , . . . ), on the fingertips, Grasp Ranking that can be deduced from the two previous items. Results Conclusion 4/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Grasp List Computation Introduction Grasp Frame Object Model Finger Workspace Method overview : Model Grasp Ranking Grasp Results Frames Grasp Quality-ordered Grasp Frames Grasp Collision + Stability Quality Score Object Model Sampling Set Computation Set Computation Grasp Set Filter Conclusion Object Surface Partitioning Finger Workspace Hand Model Approximation 5/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Grasp Frame (Hand Pose) Introduction Sampling Grasp Frame Object Model Finger Workspace Uniform sampling of frames (position + orientation) : Model Grasp • Center the frame roughly where the contact may occur. Ranking • Input ← number of positions n p , number of orientations Results n o Conclusion • Positions : uniformly sampled in the object’s axis-aligned bounding box (with a step computed to fit n p ). • The n o first elements of an incremental grid ([Yershova 2004]) 6/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Grasp Computation From Grasp Introduction Frame Grasp Frame Object Model Finger Finding the contact for a given hand pose. Workspace Model Two approaches : Grasp Ranking Results • Forward kinematics : close the fingers until contact Conclusion occur. Usually requires many collision tests. Can not find contact in loops ( e.g. mug handle). • Inverse kinematics : compute the points on object surface that are reachable by the fingers. Exact solutions is computationally very expensive. 7/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Grasp Computation From Grasp Introduction Frame Grasp Frame Object Model Finger → introduce models to quickly find a conservative Workspace Model approximation of the accessible part of the object’s surface Grasp Ranking Results → find intersection of a surface (object) and a volume Conclusion (finger workspace) • Model of object’s surface. • Model of finger workspace. 8/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Object Surface Model Introduction Grasp Frame Object Model The object surface is approximated with a point set : Finger Workspace Model • The points are obtained by uniform sampling of the Grasp object’s 3D model (triangle mesh). Ranking Results • The sampling step magnitude is chosen from the Conclusion fingertip radius. • Local information about the surface is stored with each point (surface normal and curvature). A space-partitioning tree is built upon the point set in order to have a hierarchical space partition of the points. 9/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Object Surface Points Kd-tree Introduction Grasp Frame Recursive subdivision of point set bounding-boxes. Object Model Each bounding-box is splitted in two along its larger Finger Workspace dimension until each node contain only one point. Model Grasp Ranking Results Conclusion 10/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Finger Workspace Model Introduction Grasp Frame Volumetric approximation based on a sphere hierarchy. Object Model Finger Workspace Model Grasp Ranking Results Conclusion F IG .: Schunk Anthropomorphic Hand (4 joints/3 DOFs per finger). 11/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Finger Workspace Model Introduction Construction Grasp Frame Object Model Finger Workspace Model Grasp From forward kinematics, build two point sets : Ranking • W = Points strictly inside the workspace ← sampling Results Conclusion each joint parameter over ] θ min ; θ max [ . • E = Points on the workspace boundary ← set a joint angle to θ min or θ max and sample the other ones over [ θ min ; θ max ] . 12/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Finger Workspace Model Introduction Construction Grasp Frame Object Model input : W = a set of points strictly inside the finger workspace ; Finger Workspace E = a set of points on the envelope of the finger workspace ; Model k max = the desired maximal size of the sphere decomposition ; Grasp r min = the desired minimal sphere radius ; Ranking output : S= a set of spheres S k ordered from the biggest to the smallest ; Results S = ∅ ; k = 1 ; Conclusion while k < k max do foreach p ∈ W do d ( p ) = p i ∈ ( E ∪ S ) ( � p − p i � ) ; min p best = { p ∈ W : d ( p ) = max p i ∈ W ( d ( p i )) } S k = sphere ( center = p best , radius = d ( p best )) ; S = S ∪ S k ; W = W − { p ∈ W : p ⊂ S k } ; k = k + 1 ; if d ( p best ) < r min then break ; return S ; 13/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Intersection between object Introduction surface and finger workspace Grasp Frame The two hierarchies are tested from their respective roots. Object Model This requires two elementary operations : Finger Workspace • box-sphere intersection Model • point-sphere inclusion Grasp Ranking Results Conclusion 14/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Intersection between object Introduction surface and finger workspace Grasp Frame Object Model Finger Workspace Points in the intersection are reachable but can lead to Model collisions. Grasp Ranking Results Conclusion 15/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Intersection between object Introduction surface and finger workspace Grasp Frame Object Model Finger Workspace Model foreach finger i ∈ � 1 ; nbFingers � do Grasp foreach S j ∈ � 1 ; nbSpheres � do Ranking point_set= intersect( S j , object_tree) ; Results foreach p ∈ point_set do Conclusion set_finger_config_from_IK( p ) ; collision_test( finger i , object) ; collision_test( finger i , palm ) ; collision_test( finger i , finger 1 ,..., i − 1 ) ; if no collision then → next finger ; 16/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Grasp Filtering and Ranking Introduction Grasp Frame • Grasps that do not verify force-closure are discarded. Object Model • A stability score is computed (based on [Bounab 2008]) Finger Workspace Model The above stability criterion does not guaranty the Grasp robustness of the grasp with respect to localization error. Ranking → favor contact on areas where the surface normal is less Results Conclusion varying (low curvature). Finding a trade-off between different scores. 17/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Some Results Introduction Computation time : 5 minutes on a standard PC for a list Grasp Frame with more than 100 grasps. Object Model Finger Workspace Model Grasp Ranking Results Conclusion 18/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation

Selecting a Grasp Interactively Introduction Grasp Frame Object Model • The grasp set can be browsed to select a grasp Finger adapted to the context. Workspace Model • Arm and/or base inverse kinematics are tested for each Grasp Ranking grasp of the list until a solution is found : Results Conclusion 19/ 21 IROS’10 Workshop: Grasp Planning and Task Learning by Imitation