Collaborative Grasp Planning with Multiple Object Represent - PowerPoint Presentation

Slide1

Collaborative Grasp Planning with Multiple Object Representations

Peter Brook

Matei Ciocarlie

Kaijen HsiaoSlide2

Uncertainty in Object Perception

Sensed Scene

Real SceneCollaborative Grasp Planning with Multiple Object Representations2Slide3

Uncertainty in Object Perception

Sensed Scene

Misdetected ObjectCollaborative Grasp Planning with Multiple Object Representations3Slide4

Different Object Representations

Collaborative Grasp Planning with Multiple Object Representations

4Known object grasps precomputed using GraspIt!Minimum value bounding box decomposition [Huebner et. al., ICRA 2008]

Novel object grasps from point clouds[Hsiao et al., IROS 2010]Grasp Planning via Decomposition Trees[Goldfeder et. al., ICRA 2007]Many moreSlide5

Why Choose Just One?

One object – multiple representationsPool grasps from all representations

Evaluate grasps on all representationsCombine evaluations for each grasp based on representation confidenceA good grasp on the real object should be likely to work on all likely representationsCollaborative Grasp Planning with Multiple Object Representations5Slide6

Predicting Grasp Success

Collaborative Grasp Planning with Multiple Object Representations

6A good grasp on the real object should be likely to work on all likely representationsSlide7

ROS Grasping Pipeline

Collaborative Grasp Planning with Multiple Object Representations

7Slide8

Collaborative Grasp Planning with Multiple Object Representations

Our Implementation

8Slide9

Grasp Evaluation on Object Representations

Evaluate grasp on all representations

For point cluster grasps:Heuristic quality measuresFor database objects:Evaluate in simulation (expensive)Use data-driven regressionCollaborative Grasp Planning with Multiple Object Representations9Slide10

Combining Object Representations

Relative weightings between

, the cluster representation, and , database object

 Collaborative Grasp Planning with Multiple Object Representations10

 Slide11

Object Detection

GraspIt

! energy metricCollaborative Grasp Planning with Multiple Object Representations11

Based on 892 point clouds of 44 objectsBased on 490 recorded grasps of 30 objectsModelsSlide12

Simulation experiments

Collaborative Grasp Planning with Multiple Object Representations

12Slide13

Simulation experiments

Single object on tablePoint cloud recorded from stereo camera with labeled ground truth object pose

We know mapping from simulation score to real successEvaluate grasps in simulation on true objectCollaborative Grasp Planning with Multiple Object Representations13Slide14

Simulation Results

Collaborative Grasp Planning with Multiple Object Representations

14How does collaborative planning compare to the existing planning method on known and unknown objects?Slide15

PR2 Results

Object Category

Collaborative PlannerNaïve PlannerNovel22/2518/25Database22/2521/25Collaborative Grasp Planning with Multiple Object Representations15 Single object on table Success = lift and move to side without droppingSlide16

Conclusion and Future Work

Framework for integrating different grasping algorithms and representationsEncourages detection algorithms to expose their uncertainty

Combining models allows robust graspingCollaborative Grasp Planning with Multiple Object Representations16

Add more!RepresentationsGrasp generation and evaluation algorithmsBayesian Grasp PlannerFeedback to detection algorithmsFeedback to high-level decisions about grasping vs. exploring