Combining Motion Planning and Optimization for Flexible Robot Manipulation

Jonathan Scholz and Mike Stilman

International Conference on Humanoid Robotics, 2010

COMP 790-099, Presenter: Ravikiran Janardhana

1

Problem Statement

• Design a system/algorithm to solve general manipulation tasks in natural human environments

• Involves uncertain dynamics and underspecified goals

• Service Manipulation Tasks – House Cleaning to Collaborative Factory Automation

2

Service Robots

• Challenges – Unfamiliar Objects and Abstract Goals

• Learn about objects in addition to planning interactions

• Accept broad variety of goalsEg:- Setting a table

3

Related Work

• Probabilistic Roadmaps, Rapidly Exploring Random Trees

• Model-free Reinforcement Learning

• Model-based learners i.e., learning from demonstration

4

Proposed Solution

• Task space based probabilistic planner

• Combine strengths of model based planning and reinforcement learning i.e., model-based planning with optimization

• Reaching an optimal world configuration is more important than finding the optimal way to reach it

5

Flexible Manipulation

• Determining the goal or the optimal configuration

• Finding the forward models for robot actions

• Planning to use the actions to reach the goal

6

Service Task: Setting a Table

• Consider a dinner where n guests must be given n plates and m platters must be placed at the center of the table

7

Objective Function Specification

• User can specify the goal as an abstract optimization metric

• Following are the objectives:-

– The plates should be located far from each other

– The platters should be at the center of the table

– The platters should be aligned parallel to the table

8

Objective Function Specification

• Define two sets of objects: plates P and platters Q

• Each object location is parameterized by position and orientation {x, y, θ}

• Environmental constraints – Table Dimensions

xmin ≤ x ≤ xmax; ymin ≤ y ≤ ymax;

9

Objective Function - Math• Maximize Plate distance

• Put Platters at Table Center

• Align Platters with Table

10

Objective Function - Math• Overall objective function:

• The weights α, β, γ must be specified with regard to the relative importance of the subtasks.

11

Action Model Learning

• Given state space S and actions A, probability of outcome of any action in any state is

• Probability distribution obtained by exploration.

• Compute probability models of displacement,

12

Motion Primitives

13

Forward Models

14

Models Achieved

15

Learning Forward Models - Demo

16

Motion Planner (Task Space RRT)

17

Experiments / Results

18

Experiments / Results

19

Experiments / Results - Demo

20

Experiments / Results

21

Conclusion / Future Work

• The paper presents a general framework for handling abstract tasks in object manipulation using reinforcement learning and model based planning

• Explore broader tools and domains that increase the generality of task space planning by combining planning, learning and optimization

22

Comments

• Requires tuning of parameters such as σ2ref and ɛ

which are highly task dependent

• Models can be stored for future use

• Collision detection would be complex if problem size was increased, RRT might then become deadlocked and algorithm is reduced to random search

23

Q&A

24