27
Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October 23, 2006 Trestle Project Robotics Institute Carnegie Mellon University

Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

Embed Size (px)

Citation preview

Page 1: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

Towards Proactive Replanningfor Multi-Robot Teams

Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for SpaceOctober 23, 2006 Trestle

ProjectRobotics Institute

Carnegie Mellon University

Page 2: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 2

Motivation

•Human workers:• Predict likely outcomes• Move between teamsmid-task

•Can multi-robotteams do thesame?

Page 3: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 3

What is Proactive Replanning?

• Predict problems and opportunities

•Replan before they manifest

Page 4: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 4

Idea

• Iterative repair planner•Add Proactive Replanning

• Duration Prediction• Live Task Modification

•Replan and modify active teams to:

• Forestall problems• Grasp opportunities

Page 5: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 5

Results Preview

• Stochastic domain•Metric is schedule makespan•Makespan reductions of 11-32%

Page 6: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 6

Approach Overview

•Domain•Architecture•Duration Prediction• Live Task Modification

Page 7: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 7

Domain

•Multi-agent, multi-team assembly•Goal: Minimize schedule length

Page 8: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 8

Scenario

Page 9: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 9

Architecture

Planner (ASPEN):• Centralized• Repairs & optimizes schedule• Dispatches tasks• Duration Prediction & Live Task Modification

Executive:• Manages execution of tasks• Monitors resource usage• Transmits state to planner

Behavioral:• Interfaces with hardware• Transmits state to executive• Behavioral and hardware

both simulated

Page 10: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 10

Planner

•ASPEN: Iterative repair and optimization•Duration Prediction within constraint network• Live Task Modification during repair and optimization

Page 11: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 11

Planner: Conflict Resolution

Page 12: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 12

Planner: Optimization

•Metric: schedule length•Use idle agents to:

• Start tasks on the “critical path”• Speed up executing tasks on the critical path

Page 13: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 13

Duration Prediction: Why?

Page 14: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 14

Duration Prediction: How?

• Predict remaining duration at each timestep•Replan in response•Challenge:

• Accurate predictions within resource bounds

•Current approach:• Offline simulation + lookup table

Page 15: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 15

Live Task Modification: Why?

Page 16: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 16

Live Task Modification: How?

•As part of schedule repair or optimization

• Heuristically select a new team• Subject to constraints• Currently assume instant transfers

Page 17: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 17

Live Task Modification: How?

•Challenge: search large space of teams and agents

Page 18: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 18

Experimental Results

• Scenario•Conditions•Data

Page 19: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 19

Experimental Scenario

Page 20: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 20

Experimental Approach

• 50 simulated assemblies per condition• 4 conditions

Page 21: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 21

Baseline Condition

•ASPEN•No Proactive Replanning• Each time step:

• Right-shift• Left-shift• Optimize• Repair

Page 22: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 22

Experimental Conditions

•Baseline, plus:• Prediction, or:• Live Modification, or:• Combined

Page 23: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 23

Summary of Results

Page 24: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 24

Results Details

50 runsper conditionmean (std dev)

Baseline(ASPEN)

Prediction Modification Combination

Schedule length (s)

1176.90 (343.58)

1050.14 (273.32)

820.84 (123.55)

802.76 (141.53)

Reduction in length

---- (----) 10.8% 30.3% 31.8%

Repair episodes

30.04 (36.60)

123.28 (23.06)

15.78 (8.00)

97.38 (23.86)

Useful team modifications

35.26 (6.15)

35.14 (5.46)

51.64 (6.19)

48.58 (7.58)

Page 25: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 25

Future Work

• Function approximation and Duration Prediction•Durative agent transfers•Risk management

Page 26: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 26

Summary

• Initial implementation of Proactive Replanning•Results are promising:

• Makespan reductions of up to 32%• Further work is underway

Page 27: Towards Proactive Replanning for Multi-Robot Teams Brennan Sellner and Reid Simmons 5th International Workshop on Planning and Scheduling for Space October

IWPSS 2006 - Brennan Sellner [email protected] Slide 27

Thanks!

The executive's first name was Tanner,

A shy, but proactive, replanner Who solved every trouble With a change, on the double,Which finished the job in fine manner.