Picking Up the Pieces: Grasp Planning via Decomposition Trees Corey Goldfeder, Peter K. Allen, Claire Lackner, Raphael Pelosoff.

Picking Up the Pieces: Grasp Planning via

Decomposition Trees

Corey Goldfeder, Peter K. Allen, Claire Lackner, Raphael

Pelosoff

Grasp Synthesis High dimensional, nonlinear space

configuration space = joints + pose grasp quality is not smooth

Difficult to model analytically Must account for dynamics, soft

contacts, non-fingertip contacts, material properties

Many constraints Obstacles, hand kinematics and scale

Our approach Simulation based grasp synthesis

has many advantages

Space of all grasps is too large to explore fully in simulation

We want a subspace that contains many good grasps

GraspIt! Grasp simulator for both robotic

and human hands

Includes kinematics,dynamics

Real time 3D visualization

Efficiently computesgrasp quality

Graspit!: A Versatile Simulator for Robotic Grasping, IEEE Robotics and Automation Magazine, 11.4

Grasping By Parts Automatic Grasp Planning

Using Shape Primitives -Miller et. al.

Superquadrics Simple volumetric primitive

Small parameter space (11 dimensions)

Preserves approximate normals

Segmentation and Superquadric Modeling of 3D Objects

- Chevalier, Jaillet, Baskurt

We added nearest neighbor pruning to reduce complexity by a factor of n

Split-Merge Decomposition

Decomposition Trees

A model…

8 levels of decomposition

……the decomposition

tree

Decomposition Trees Building a tree from the bottom up

Pairwise merge of parts with least error

How Many Parts? Use an error threshold?

Problem: large superquadrics can swallow important features, like handles, without much error

Solution: fixed number of parts decompose all objects to n

superquadrics n is chosen experimentally for a

given hand

Planner Overview Decompose into tree with n leaves

Plan grasps on superquadrics using entire tree, not just leaves

Simulate candidates on actual geometry, using GraspIt!

Rank results by grasp quality

Results Planned multiple stable grasps for

all our test objects

Results Works even for objects difficult to

represent with superquadrics

Difficulties Assumes knowledge of object

geometry

Superquadric decomposition is slow

Grasping a single part is done heuristically

Cannot plan candidates on parts from different branches of the tree

Do Trees Help?

Without treesWith trees

Without a tree, some good grasps

With a tree, many good grasps if a grasp is

unsuitable, another good grasp can be substituted

Contributions Fully automatic implementation of

grasping-by-parts

Abstracts away fine features

Allows multiple parts to be planned on as a group

Future Work Incorporate existing SVM planner

for individual superquadrics

Speed up decomposition

Questions?