A Probabilistic Approach to Grasp Planning - ROS.org · MotivationPlanner...

Motivation Planner Overview Evaluation Conclusion

A Probabilistic Approach toGrasp Planning

Peter Brook

Department of Computer ScienceUniversity of Washington

Willow Garage, Inc

September 23, 2010


What is grasp planning?

Given an object:I Find a suitable pose for the gripper - 6D search spaceI Find a suitable configuration of the hand joints - nD search

space

A Probabilistic Approach to Grasp PlanningPeter Brook


Why is Grasp Planning Easy for Humans?



Why is Grasp Planning Hard for Robots?



Uncertainty in grasp planning

I Grasp planning requires knowledge of:I a gripperI an object

I These are observed via sensors




Real scene Sensed scene




Is incomplete or noisy world information a problem?




Yes.

Sensed scene Object detection errorA Probabilistic Approach to Grasp PlanningPeter Brook


Video

Play Video



ROS Grasping Pipeline

3D Perception

Grasp Planningfor unknown objects

Collision Map Generation

Scene Interpretor

Motion Planning

Object Model Registration

Object Model Database

Grasp Planningfor known objects

Grasp Selection

Grasp Execution

Grasp Execution

Tactile Feedback

Grasp Planning



Handling Object Detection Uncertainty

Key PointsI Every guess about the

object is a representationI Choose grasps which work

well on availablerepresentations



How do we Know if a Grasp is Good?

I evaluate grasp on allrepresentations

For point cluster grasps:I Heuristic quality measures

For database objects:I Evaluate grasp in

simulation (expensive)

I Use data-driven regressionfor evaluation



How do we Know if a Grasp is Good?

I evaluate grasp on allrepresentations

For point cluster grasps:I Heuristic quality measures

For database objects:I Evaluate grasp in

simulation (expensive)I Use data-driven regression

for evaluation



Gripper Pose Uncertainty

I Bonus: regression helpswith gripper poseuncertainty



Tuning the System



Reasoning About Uncertainty Helps

0.0 0.2 0.4 0.6 0.8 1.0

Planner-Estimated Probability

0.60

0.65

0.70

0.75

0.80

0.85

0.90

0.95

1.00

Perc

enta

ge

of

’Good

’G

rasp

s

Naive Planner

Naive Planner (leave-one-out)



Reasoning About Uncertainty Helps

0.0 0.2 0.4 0.6 0.8 1.0

Planner-Estimated Probability

0.60

0.65

0.70

0.75

0.80

0.85

0.90

0.95

1.00

Perc

enta

ge

of

’Good

’G

rasp

s

Naive Planner

Naive Planner (leave-one-out)

Probabilistic Planner

Probabilistic Planner (leave-one-out)



Experimental Success Results

prob. planner naive planner

leave-one-out 22/25 18/25regular detection 22/25 21/25

Table: Number of objects successfully grasped and lifted on the PR2robot.



Video

Play Video



Potential for Extension

I More representationsI Primitives (spheres, cylinders, boxes, etc.)I Primitive hierarchies

I More detectors (recognition pipeline)I Don’t hide the uncertaintyI Detector feedback: similar objects should be grasped

similarlyI More objects

I few objects: no generalizationI 50 objects: some generalizationI 1000s of objects: complete generalization?



Can we Trust Our Sensors?



Online Tuning



Code Availability

I Planner code: probabilistic_grasp_plannerI Documentation:

http://ros.org/wiki/probabilistic_grasp_planner

I Integrated with trunk of grasping pipelineI Will be released with the 0.3 release of

object_manipulation


http://ros.org/wiki/probabilistic_grasp_planner


Thanks!

Matei and Kaijen are awesome!Questions?


A Probabilistic Approach to Grasp Planning - ROS.org · MotivationPlanner...

Documents

Transcript of A Probabilistic Approach to Grasp Planning - ROS.org · MotivationPlanner...