Implementation of Multi-Goal Motion Planning Under Uncertainty on a Mobile Robot
ICRoM 2017
November 20, 2017 Ali Noormohammadi Asl
Hamid D. Taghirad Amirhossein Tamjidi
Dept. of Electrical Engineering K. N. Toosi University of Technology Advanced Robotics and Automated Systems (ARAS)
16 Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase Online Phase
Conclusion
Advanced Robotics and Automated Systems (ARAS)
Introduction Problem Statement
Multi-goal Motion Planning POMDP
Problem Formulation TSP-FIRM
TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase
Online Phase Conclusion
Introduction 2 Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase Online Phase
Conclusion
Introduction
Motion planning under uncertainty
I Finding a collision-free motion for the robot system from an initial configuration to goal configuration.
I The uncertainty is closely intertwined with most robotic applications.
I Uncertainy emanates from three main sources:
I Model (motion) uncertainty
I Observation (sensor) uncertainty
I Environment (map) uncertainty
I Ignoring uncertainty in planning and decision making may well cause improper results and even failure.
16 Introduction 2 Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase Online Phase
Conclusion
Advanced Robotics and Automated Systems (ARAS)
Motion planning under uncertainty
I Finding a collision-free motion for the robot system from an initial configuration to goal configuration.
I The uncertainty is closely intertwined with most robotic applications.
I Uncertainy emanates from three main sources:
I Model (motion) uncertainty
I Observation (sensor) uncertainty
I Environment (map) uncertainty
I Ignoring uncertainty in planning and decision making may well cause improper results and even failure.
Introduction Problem Statement
3 Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase Online Phase
Conclusion
Problem Statement
Multi-goal Motion Planning
Multi-goal Motion Planning
I Finding a policy for the robot to traverse through a number of goal points.
1 Start
2
3
4
5
6
7
16 Introduction Problem Statement
3 Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase Online Phase
Conclusion
Advanced Robotics and Automated Systems (ARAS)
Multi-goal Motion Planning
Multi-goal Motion Planning
I Finding a policy for the robot to traverse through a number of goal points.
I Simple TSP: A primitive solution in the absence of uncertainty.
1 Start
2
3
4
5
6
7
Introduction Problem Statement
3 Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase Online Phase
Conclusion
Problem Statement
Multi-goal Motion Planning
Multi-goal Motion Planning
I Finding a policy for the robot to traverse through a number of goal points.
I Simple TSP: A primitive solution in the absence of uncertainty.
I The TSP problem should be solved in belief space.
I The path between each two goal points is in not the direct line connecting them.
I The costs are not deterministic.
I The graph is not symmetric.
16 Introduction Problem Statement
Multi-goal Motion Planning POMDP 4 Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase Online Phase
Conclusion
Advanced Robotics and Automated Systems (ARAS)
POMDP
Partially Observable Markov Decision Processes
I Motion planning under uncertainty is an instance of the problem of sequential decision making under uncertainty.
I Partially Observable Markov Decision Processes (POMDP) is a general framework for sequential decision making under uncertainty.
J(·) =min
Π
∞
X
k=0
E[c(bk, π(bk))]
π∗=argmin
Π
∞
X
k=0
E[c(bk, π(bk))]
s.t. bk+1=τ(bk, π(bk),zk), zk ∼p(zk |xk)
Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation 5 TSP-FIRM
TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase Online Phase
Conclusion
Problem Statement
Problem Formulation
Problem Formulation
I Multi-goal motion planning formulation using TSP and POMDP frameworks.
{p,Π}min
N
X
i=1 N
X
j=1
pij
∞
X
k=0
E
cj bik, π bki
s.t.
N
X
j=1
pij=1 (i6=j,i∈V)
N
X
i=1
pij=1 (i6=j,j∈V)
Xpij≤ |s| −1 (s⊂V , 2≤ |s| ≤N−2) pij∈ {0,1} (i,j)∈A
bk+1=τ(bk, π(bk),zk)
I This optimization problem is intractable to solve.
16 Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation 5 TSP-FIRM
TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase Online Phase
Conclusion
Advanced Robotics and Automated Systems (ARAS)
Problem Formulation
Problem Formulation
I Multi-goal motion planning formulation using TSP and POMDP frameworks.
{p,Π}min
N
X
i=1 N
X
j=1
pij
∞
X
k=0
E
cj bik, π bki
s.t.
N
X
j=1
pij=1 (i6=j,i∈V)
N
X
i=1
pij=1 (i6=j,j∈V)
Xpij≤ |s| −1 (s⊂V , 2≤ |s| ≤N−2) pij∈ {0,1} (i,j)∈A
bk+1=τ(bk, π(bk),zk)
I This optimization problem is intractable to solve.
Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM 6 TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase Online Phase
Conclusion
TSP-FIRM
I Feedback based Information RoadMap is a multi–query graph based algorithm for path planning under uncertainty.
I FIRM helps to simplify the complex POMDP problem into a tractable MDP problem.
I We use FIRM to solve the TSP optimization problem in the belief space.
16 Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM 7 TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase Online Phase
Conclusion
Advanced Robotics and Automated Systems (ARAS)
TSP-FIRM Graph
TSP-FIRM Graph
1. Forming a Probabilistic RoadMap (PRM) including the goal points and the sampled nodes.
2. design a stabilizer (node controller) for each node.
I each TSP-FIRM node is a small region
B={b:kb−b0k ≤}around the sampled beliefb0. 3. Designing TSP-FIRM graph edges (local controllers).
I Each edge is a concatenation of the node controller and the edge controller.
Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM 7 TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase Online Phase
Conclusion
TSP-FIRM
TSP-FIRM Graph
TSP-FIRM Graph
I A graph with the set of nodes and the set of edges is obtained.
I Motion planning over the entire belief space is reduced to planning over the TSP-FIRM graph.
1 2
3
4
5
6
7
16 Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph
8 Obstacles in TSP-FIRM TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase Online Phase
Conclusion
Advanced Robotics and Automated Systems (ARAS)
Obstacles in TSP-FIRM
Obstacles
There are three types of obstacles.
1. Completely known obstacles 2. Areas which are potentially obstacle 3. Completely unknown obstacles
One-step-cost
Cig(B,u) =0 if B∈Bgoali
or (F happens) and(B∈/ Fsus)
Cig(B,u)≥β > ε >0 if (B∈Fsus) Cig(B,u)≥ε >0 if otherwise
Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements 9 Experimental Results
Implementation Details Offline Phase Online Phase
Conclusion
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
TSP-FIRM Nodes
I Designing a stationary Kalman filter (SKF) for each sampled node.
I Constructing TSP-FIRM node with the center bcj ≡
vj,Psj .
Bj =n
b≡(x,P) : x −vj
< δ1,
P−Psj
m< δ2o
16 Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements 10 Experimental Results
Implementation Details Offline Phase Online Phase
Conclusion
Advanced Robotics and Automated Systems (ARAS)
TSP-FIRM Elements
Local Controller
I Designing a Kalman filter for the robot position estimation.
I Designing a switching controller for the stabilizer (node controller).
I Designing an LQG controller for the edge controller to drive the robot from a node to another one.
Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements 11 Experimental Results
Implementation Details Offline Phase Online Phase
Conclusion
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Transition Cost and Probability
Using the sequential Monte Carlo method for computing transition costs and probabilities.
I Estimation accuracy,Φij =E hPT
k=1tr
wPkiji .
I Stopping time of local controller,Tˆij =E Tij
.
I The time robot moves in the high-risk area,Tˆobsij =E Tij
.
C Bi, µij
=ξ1Φij+ξ2Tˆij+ξ3Tˆobsij
16 Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements 11 Experimental Results
Implementation Details Offline Phase Online Phase
Conclusion
Advanced Robotics and Automated Systems (ARAS)
TSP-FIRM Elements
Transition Cost and Probability
Using the sequential Monte Carlo method for computing transition costs and probabilities.
I Estimation accuracy,Φij =E hPT
k=1tr
wPkiji .
I Stopping time of local controller,Tˆij =E Tij
.
I The time robot moves in the high-risk area,Tˆobsij =E Tij
.
C Bi, µij
=ξ1Φij+ξ2Tˆij+ξ3Tˆobsij
Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details 12 Offline Phase Online Phase
Conclusion
Experimental Results
Implementation Details
Motion Model
Differential wheeled mobile robot: MELON
Xk+1=f(Xk,uk,wk) =
xk+ (Vk+nv)δtcosθk
yk+ (Vk+nv)δtsinθk
θk+ (wk+nw)δt
Qk=diag
ηvVk+σVb2
,(ηwwk+σbw)2
16 Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details 12 Offline Phase Online Phase
Conclusion
Advanced Robotics and Automated Systems (ARAS)
Implementation Details
Motion Model
Differential wheeled mobile robot: MELON
Xk+1=f(Xk,uk,wk) =
xk+ (Vk+nv)δtcosθk yk+ (Vk+nv)δtsinθk
θk+ (wk+nw)δt
Qk=diag
ηvVk+σVb2
,(ηwwk+σbw)2
Sensor Model
Black and white patterns: ArUco markers
jzk= jdk
,atan2 jd2k,jd1k
−θT
+jv,jv∼ N 0,jR
jRk=diag ηrd
jdk
+ηrφ|φk|+σbr2
, ηθd
jdk
+ηθφ|φk|+σθb2
Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details 13 Offline Phase Online Phase
Conclusion
Experimental Results
Implementation Details
Map
Second floor of the Electrical Engineering department at K. N.
Toosi University of Technology.
16 Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details
14 Offline Phase Online Phase
Conclusion
Advanced Robotics and Automated Systems (ARAS)
Offline Phase
Offline Phase
I Choosing goal points
I Generating graph
Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details
14 Offline Phase Online Phase
Conclusion
Experimental Results
Offline Phase
Offline Phase
I Choosing goal points
I Generating graph
I Calculating transition cost, transition probability and failure probability
I Finding the best way between each two goal points
I Solving TSP:[1,2,3,4,5,6,7,1]
[1,10,9,8,2,11,12,3,14,5,15,6,22,7,21,17,16,4,13,3,12, 11,2,8,9,10,1]
16 Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase
15 Online Phase Conclusion
Advanced Robotics and Automated Systems (ARAS)
Online Phase
Online Phase
The robot starts executing the offline obtained policyBUT
Challenges in the Online Phase
I Change in map
I Deviation from planned path
I Observation loss
I Robot Kidnapping
An online replanning algorithm is proposed to update map, graph, and policy
Introduction Problem Statement
Multi-goal Motion Planning POMDP Problem Formulation
TSP-FIRM TSP-FIRM Graph Obstacles in TSP-FIRM
TSP-FIRM for a Nonholonomic Robot
TSP-FIRM Elements
Experimental Results Implementation Details Offline Phase Online Phase
Conclusion 16
Conclusion
I Modeling multi-goal motion planning problem as an asymmetric traveling salesman problem (ATSP) in the belief space.
I Presenting algorithms for the offline generation of TSP-FIRM graph and online execution.
I Generilizing the algorithms for nonholonomic mobile robots.
I Implementation on a physical mobile robot in a real environment.