decentralized path planning for multi-agent teams in ... filevishnu r. desarajuand jonathan p. how...

23
VISHNU R. DESARAJU AND JONATHAN P. HOW DEPARTMENT OF AERONAUTICS AND ASTRONAUTICS MASSACHUSETTS INSTITUTE OF TECHNOLOGY ICRA 2011 AUTONOMOUS ROBOTS 2012 PRESENTED BY: RAJ DASGUPTA Decentralized Path Planning for Multi-Agent Teams in Complex Environments using Rapidly-exploring Random Trees

Upload: hoangnhan

Post on 12-Aug-2019

214 views

Category:

Documents


0 download

TRANSCRIPT

VISHNU R . DESARAJU AND JONATHAN P . HOW

DEPARTMENT OF AERONAUTICS AND ASTRONAUTICS

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

ICRA 2011

AUTONOMOUS ROBOTS 2012

PRESENTED BY: RAJ DASGUPTA

Decentralized Path Planning for Multi-Agent Teams in Complex Environments using Rapidly-exploring Random Trees

Overview

� Path planning in multi-robot (agent) system

� Sampling-based planner

� New algorithm called DMA-RRT proposed

� Includes coordination strategy based on token ring passing between robots

� Extension to basic algorithm called Cooperative DMA-RRT allows agents to update each other’s plans

� Verified experimentally through simulations and on iRobot Create platform

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

2

Problem Addressed

� Plan a collision free path for multiple agents

� Constraints

� Complex environment

� Agents (called a team) should not collide with each other

� Static or linear paths cannot be used - requires paths to be dynamically updated

� Centralized planning cannot be used

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

3

Related Work: Robot Path Planning

� Dynamic path planning using constraint solvers scale poorly in number of constraints

� Mixed Integer Linear Programming (MILP)

� Model Predictive Control (MPC)

� Potential field based planning – works for simple environments only

� Sampling-based methods

� Probabilistic Roadmap Planners (PRM)

� Rapidly-exploring Random Trees (RRT)

� Extension called closed loop RRT (CL-RRT) used as basis of work

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

4

Related Work: Robot Coordination

� Agents must reach consensus on updated path before moving (Scerri et al.)� Synchronization of plans causes delay

� Agents reserve regions of map that they will move into and must reach consensus about changes to those regions (Purvin et al.)� Asynchronous planning allowed, but some tradeoff due to working with large areas instead of paths

� Reachability-based collision avoidance considers non-cooperating agents (worst case scenario)

� Decentralized MPC (Trodden and Richards): Agents allowed to update plans sequentially in pre-determined order - solves synchronization issues; fixed order makes it inefficient

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

5

CL-RRT Algorithm

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

6

� Generates a set of feasible paths � Satisfying current path constraints

� Selects least cost path from set

� Not guaranteed to be optimal

� Fast (faster than comparable path planner RRT*)

� Verified successfully on robots in DARPA Grand Challenge 2007

� Three phases� Tree expansion

� Execution loop

� Lazy check for feasibility

CL-RRT Algorithm

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

7

Multi-agent CL-RRT

� Naïve approach: Every agent plans its path subject to constraints imposed by other agents� Problem: Inefficiency

� A1 calculates its plan and sends to A2 as set of constraints

� A2 starts calculating its plan using A1’s constraints

� A1 updates its plan and corresponding set of constraints

� A2 finishes plan with old set of A1’s constraints; sees A1 has new constraints; replans…

� More complicated and inefficient with more agents

� CL-RRT can fail (cause collisions) when used by multiple agents with inter-dependent path constraints plan simultaneously

� Needs coordination protocol

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

8

Coordination Protocol: Merit-based Token Passing

� Basis: (Trodden and Richards, ACC 2006)� Sequential order imposed among agents � In each iteration only one agent can update its plan

� Merit-based Token Passing� Remove sequential order by introducing a token� Agent that has token can update its plan� Each agent without token calculates metric called Potential Path Improvement (PPI) � PPI = difference in cost between best path in RRT and current path

� Broadcasts PPI to all other agents� In next round token goes to agent with highest PPI from last round; ties broken at random

� Advantages: � Agent that benefits most from re-planning is given next chance to replan� Agents that can quickly find path to goal get token first without waiting in a sequential order

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

9

Decentralized Multi-Agent (DMA) RRT

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

10

DMA-RRT: Path Constraint Satisfaction

Main idea: DMA-RRT does not introduce any new path conflicts; proven by induction

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

11

DMA-RRT: Merit-based token passing

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

12

DMA-RRT vs. Cooperative DMA-RRT

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

13

Cooperative DMA-RRT Algorithm: 1

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

14

Cooperative DMA-RRT Algorithm: 2

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

15

Experimental Results

� Simulations:

� Each agent simulated on 1 desktop (2.13 GHz, 1 GB RAM)

� 10 desktops connected by LAN

� Two wheel skid steer robot model

� Path cost=Travel time along path

� Physical robot experiments

� iRobot Create

� Pure pursuit controller used to determine robot velocity

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

16

Simulation Results: Open Map, 10 Agents

� No obstacle environment with 10 goals (waypoints)

� Agents cycle through goals for 10 mins (results avg. over 12 trials)

� Fig. 2(b) shows agent 1’s view of other agents trajectories (obstacles) with round robin sched.

� Legends:� Blue circle/red arrow : agent 1� Yellow circle/blue arrow: other agents� Magenta dots: waypoints for agent 1’s path

� Red: current position of other agent� Green: future position (+10 sec.) other agent

� Red-green: time parameterized obstacles presented by other agents’ paths

Round-robin scheduling: 12.5 goals/agent in 10 mins.Merit-based scheduling: 15.1 goals /agent in 10 minsImprovement because agent that finds a shorter (better) path to goal does not have to wait for all other agents to finish their turn before it gets token and can replan2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

17

Simulation Results: Open Map – Replan Times

� Round-robin gives regular replan start times (uniformly spaced dots in Fig. 3 (a))

� Merit-based gives task-driven replan start times (non-uniformly spaced dots in Fig. 3 (b))

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

18

Experiment 2: 4 agents with obstacles

� Sim. Time: 10 mins� Round robin: 11.9 goals/agent (in 10 mins)

� Merit-based: 11.9 goals/agent (in 10 mins)� Too few agents, little delay in getting token – same wait time as round robin

� Cooperative: � Stop nodes every 4 sec. along trajectory

� 13.0 goals/agent� Emergency stops at or near openings of narrow passage between obstacles – helps agents replan

� Cooperation strategy prevents other agents from blocking path found by an agent through narrow passages

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

19

Experiment 3: Open map, physical robots

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

20

Experiment 4: Emergency Stop with Cooperative DMA-RRT

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

21

Experiment 3: Collision Avoidance

� 5 min run on open map� Minimum inter-agent separation allowed = 0.3 m� Satisfied by both algorithms as per Tables I and II

� Implies both algorithms achieve collision avoidance

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

22

Conclusion

� Main contribution: Merit-based token passing integrated with CL-RRT planner

� Future work

� Communication (range) constraints

� Dynamic obstacles

� Larger simulations/experiments

2/4/2013CMANTIC Lab Reading Group: Raj Dasgupta

23