A Fast Approach in Generating High Quality Grasps Watcharapol - PowerPoint PPT Presentation

A Fast Approach in Generating High Quality Grasps Watcharapol Watcharawisetkul Adviser: Asst. Prof. Nattee Niparnan, Ph.D. 1

Outline ● Introduction ● Background Knowledge ● Our Approach ● Evaluations ● Conclusion 2

Outline ● Introduction ● Background Knowledge ● Our Approach ● Evaluations ● Conclusion 3

dvice.com darpa.mil brown.edu Robotic grasping 4

suin.mx bloggaida.blogspot.com itcentralstation.com updatedtrends.com 5

Grasping pipeline 1 2 3 4 Object Grasp Grasp Grasp perception synthesis planning execution 6

Focus on 1 2 3 4 Object Grasp Grasp Grasp perception synthesis planning execution 7

Contribution ● Propose grasp synthesis algorithm ○ Calculate a large number of high quality secure grasps in short time Input Output ● Position and normal of ● List of secure grasps object surface points 8

Outline ● Introduction ● Background Knowledge ● Our Approach ● Evaluations ● Conclusion 9

Finger model ● Hard finger with friction ○ Can exerts only pure force lie in a circular friction cone 10

Wrench ● Force ( ) ○ 3D ● Torque ( τ ) · ○ 3D T τ T ] T ● Wrench [ ○ Concatenation of force and torque ○ 6D 11

Secure grasp ● Force closure ○ Ability to resist any external disturbance ○ Contact wrenches generated by fingers positively span entire wrench space ✔ ✘ 12

Number of fingers ● X. Markenscoff et al. , “The Geometry of Grasping,” 1990. ○ Sufficient no. of fingers to archived force closure without friction with friction 2D 4 3 3D 7 4 13

Outline ● Introduction ● Background Knowledge ● Our Approach ● Evaluations ● Conclusion 14

Our Approach ● Objective ○ Propose grasp synthesis algorithm that can calculate a large number of high quality grasps in short time ● Randomized algorithm bound with time limit ● Concurrent grasp 15

Concurrent grasp Concurrent grasps ⊂ Force closure grasps Concurrent point C1. There exist lines in each friction cone that intersect in a single point C2. The vectors parallel to these lines and pointing inward positively span force space 16

Finding concurrent grasp algorithm Input : S = object surface points frictionHalfAngle timeLimit Output : G ans = list of concurrent grasp 1:while usedTime < timeLimit 2: randomly pick concurrent point p 3: M ← { s | s ∈ S, p in DFC s } 4: V ← { v m | v m = inward( p-m ) ∀ m ∈ M } 5: G ← {( a,b,c,d )| v a ,v b ,v c ,v d ∈ V, v a ,v b ,v c ,v d positively span ℝ 3 } 6: G ans ← G ans ∪ G 17

Finding concurrent grasp algorithm Input : S = object surface points frictionHalfAngle timeLimit Output : G ans = list of concurrent grasp 1:while usedTime < timeLimit 2: randomly pick concurrent point p 3: M ← { s | s ∈ S, p in DFC s } 4: V ← { v m | v m = inward( p-m ) ∀ m ∈ M } 5: G ← {( a,b,c,d )| v a ,v b ,v c ,v d ∈ V, v a ,v b ,v c ,v d positively span ℝ 3 } 6: G ans ← G ans ∪ G 18

Finding concurrent grasp algorithm Input : S = object surface points frictionHalfAngle timeLimit Output : G ans = list of concurrent grasp 1:while usedTime < timeLimit 2: randomly pick concurrent point p 3: M ← { s | s ∈ S, p in DFC s } 4: V ← { v m | v m = inward( p-m ) ∀ m ∈ M } 5: G ← {( a,b,c,d )| v a ,v b ,v c ,v d ∈ V, v a ,v b ,v c ,v d positively span ℝ 3 } 6: G ans ← G ans ∪ G 19

Finding concurrent grasp algorithm Input : S = object surface points frictionHalfAngle timeLimit Output : G ans = list of concurrent grasp 1:while usedTime < timeLimit 2: randomly pick concurrent point p 3: M ← { s | s ∈ S, p in DFC s } 4: V ← { v m | v m = inward( p-m ) ∀ m ∈ M } 5: G ← {( a,b,c,d )| v a ,v b ,v c ,v d ∈ V, v a ,v b ,v c ,v d positively span ℝ 3 } 6: G ans ← G ans ∪ G 20

Finding concurrent grasp algorithm Input : S = object surface points C1. There exist lines in each frictionHalfAngle friction cone that intersect timeLimit in a single point Output : G ans = list of concurrent grasp 1:while usedTime < timeLimit 2: randomly pick concurrent point p 3: M ← { s | s ∈ S, p in DFC s } 4: V ← { v m | v m = inward( p-m ) ∀ m ∈ M } 5: G ← {( a,b,c,d )| v a ,v b ,v c ,v d ∈ V, v a ,v b ,v c ,v d positively span ℝ 3 } 6: G ans ← G ans ∪ G 21

Finding concurrent grasp algorithm Input : S = object surface points frictionHalfAngle timeLimit Output : G ans = list of concurrent grasp 1:while usedTime < timeLimit 2: randomly pick concurrent point p 3: M ← { s | s ∈ S, p in DFC s } 4: V ← { v m | v m = inward( p-m ) ∀ m ∈ M } 5: G ← {( a,b,c,d )| v a ,v b ,v c ,v d ∈ V, v a ,v b ,v c ,v d positively span ℝ 3 } 6: G ans ← G ans ∪ G 22

Finding concurrent grasp algorithm Input : S = object surface points frictionHalfAngle timeLimit Output : G ans = list of concurrent grasp 1:while usedTime < timeLimit 2: randomly pick concurrent point p 3: M ← { s | s ∈ S, p in DFC s } C2. The vectors parallel to 4: V ← { v m | v m = inward( p-m ) ∀ m ∈ M } these lines and pointing 5: G ← {( a,b,c,d )| v a ,v b ,v c ,v d ∈ V, inward positively span force v a ,v b ,v c ,v d positively span ℝ 3 } space 6: G ans ← G ans ∪ G 23

Finding concurrent grasp algorithm Input : S = object surface points frictionHalfAngle timeLimit Output : G ans = list of concurrent grasp 1:while usedTime < timeLimit 2: randomly pick concurrent point p 3: M ← { s | s ∈ S, p in DFC s } C2. The vectors parallel to 4: V ← { v m | v m = inward( p-m ) ∀ m ∈ M } these lines and pointing 5: G ← {( a,b,c,d )| v a ,v b ,v c ,v d ∈ V, inward positively span force v a ,v b ,v c ,v d positively span ℝ 3 } space 6: G ans ← G ans ∪ G 24

Finding concurrent grasp algorithm Input : S = object surface points frictionHalfAngle timeLimit Output : G ans = list of concurrent grasp 1:while usedTime < timeLimit 2: randomly pick concurrent point p 3: M ← { s | s ∈ S, p in DFC s } 4: V ← { v m | v m = inward( p-m ) ∀ m ∈ M } 5: G ← {( a,b,c,d )| v a ,v b ,v c ,v d ∈ V, v a ,v b ,v c ,v d positively span ℝ 3 } 6: G ans ← G ans ∪ G 25

Finding concurrent grasp algorithm Input : S = object surface points frictionHalfAngle timeLimit Output : G ans = list of concurrent grasp 1:while usedTime < timeLimit 2: randomly pick concurrent point p 3: M ← { s | s ∈ S, p in DFC s } 4: V ← { v m | v m = inward( p-m ) ∀ m ∈ M } 5: G ← {( a,b,c,d )| v a ,v b ,v c ,v d ∈ V, v a ,v b ,v c ,v d positively span ℝ 3 } 6: G ans ← G ans ∪ G 26

Time complexity Input : S = object surface points frictionHalfAngle ● N. Niparnan et al. , 2006 timeLimit ○ Force-closure grasps Output : G ans = list of concurrent grasp ○ 2D objects ○ Frictionless 1:while usedTime < timeLimit ○ 4 fingers 2: randomly pick concurrent point p 3: M ← { s | s ∈ S, p in DFC s } 4: V ← { v m | v m = inward( p-m ) ∀ m ∈ M } 5: G ← {( a,b,c,d )| v a ,v b ,v c ,v d ∈ V, O (|V| 3 log|V|+K) v a ,v b ,v c ,v d positively span ℝ 3 } 6: G ans ← G ans ∪ G 27

Concurrent points Input : S = object surface points Different point lead to frictionHalfAngle difference result timeLimit Output : G ans = list of concurrent grasp 1:while usedTime < timeLimit 2: randomly pick concurrent point p 3: M ← { s | s ∈ S, p in DFC s } 4: V ← { v m | v m = inward( p-m ) ∀ m ∈ M } 5: G ← {( a,b,c,d )| v a ,v b ,v c ,v d ∈ V, v a ,v b ,v c ,v d positively span ℝ 3 } 6: G ans ← G ans ∪ G 28

Concurrent points (1) ● "Uniform" ○ Randomly select concurrent points from a uniform distribution within axis- aligned minimum bounding box ○ Use as baseline 29

Concurrent points (2) ● "NearCM" ○ Randomly select concurrent points near object's center of mass ○ Normal distribution ● Grasp quality ● ○ Ponce et al. 1997, Ding et al. 2001 ○ Distance between the centroid and the center of mass 30

Concurrent points (3) ● "MedialPoint" ○ Select concurrent points from object's medial axis ● Medial axis ○ Centers of spheres which touch the object's surface at two or more points 31

3D medial axis approximation ● Ma et al. 2011 ○ Input ■ Position and normal of object surface points ○ Output ■ Medial point corresponding to each object surface point 32

Outline ● Introduction ● Background Knowledge ● Our Approach ● Evaluations ● Conclusion 33

Evaluation objective ● Propose grasp synthesis algorithm ○ No. generated grasps ○ Grasp quality 34

Competitor Our algorithm ● "Uniform" ○ Select concurrent points from a uniform distribution ● "NearCM" ○ Select concurrent points near cm of the object ● "MedialPoint" ○ Select concurrent points from object's medial axis 35

Competitor (2) ● "Random4SP" ○ C. Borst et al. , “ Grasping the dice by dicing the grasp,” 2003. ○ Find force closure grasp 1: while usedTime < timeLimit 2: Random 4 surface points (N. Niparnan et al. 2007) 3: Apply heuristic filter (Y. Zheng et al. 2009) 4: Perform force closure test 36