A Practical, Decision-theoretic Approach to Multi-robot Mapping and Exploration

1
A Practical, Decision-theoretic Approach to
Multi-robot Mapping and Exploration
?? ??? ???
????? ??? ??????
???? ???? ?????? ??? ????? ??? ???? ????
????
????? ????? 04/09/1385
2

• Introduction
• Dynamic Coordination Architecture
• Design-theoretic Coordination
• Partial Map Localization
• Experiments
• Conclusions and Future Research

3
Introduction

• This approach uses an adapted version of particle
  filters
• The risk of false-positive map matches is avoided
  by verifying match hypotheses using a rendezvous
  approach
• Efficient exploration of an unknown environment
• Exploration and map building for large teams of
  robots
• Limited communication between robots
• No assumptions about relative start locations of
  the robots
• Dynamic assignment of processing tasks

4
Dynamic Coordination Architecture
5
Decision-theoretic Coordination
6

• ? denote an assignment that determines which
  robot should move to which target
• Cost If the target is a frontier then the cost
  is given by the minimum cost path from the
  robots position

7

• Utilities For simplicity, we assume that all
  robots have the same exploration capabilities,
  i.e. the utility only depend on the type of
  target and not the robot

8
Partial Map Localization
9
(No Transcript)
10
Implementation as particle filter
11
Partial resampling

• In the basic particle filter, all samples are
  frequently resampled based on their accumulated
  weights
• Unfortunately, in our context, such a resampling
  does not work since the weights of the samples
  may differ by orders of magnitude
• So, samples are weighted by
• As a result of this resampling procedure, all
  samples do not have the same weights, but carry
  the non-resampled weights with them into the next
  iteration.
• A very important advantage of this approach is
  that samples outside the partial map are not
  deleted but tracked until they re-enter the map

12
(No Transcript)
13

• After only short overlap, the best match is not
  yet correct. The summed probability of all
  samples inside the map is .497

14

• The robot exits the map, but already determined
  the correct match.
• Now, the probability of being inside the map
  dropped to 0.00037 , since all high-weight
  samples just exited the map

15

• After moving about 50m outside the map, the robot
  returned and the match is correct (probability of
  this match is .799

16
(No Transcript)
17
(No Transcript)
18
(No Transcript)
19
(No Transcript)
20

• Robots A and B start at unknown locations and
  explore independently.
• The trajectories of A and B are shown as dotted
  and solid lines, respectively.
• After some time, the robots reach positions Ia
  and Ib, and A estimates Bs location in its map.
• The corresponding maps are shown in Ia and Ib.
• The overlap between the two maps is not
  sufficient to create a hypothesis with
  probability above. Both robots keep exploring
  until, at positions IIa and IIb, A finds a very
  likely hypothesis for Bs position.
• Both robots move to the meet point and verify the
  hypothesis. The maps are merged and the robots
  start coordinated exploration.
• A moves to the left and B first moves into the
  small hallway in the lower part.

21
??????
?