CSE-571 Probabilis1c Robo1cs Exploration Map an unknown area - - PowerPoint PPT Presentation

1

SA-1

CSE-571 Probabilis1c Robo1cs

Exploration

SA-1

► Map an unknown area ► Search for an “object of value” ► Set up a surveillance network ► Track any intruders Sponsored by DARPA-SDR, NSF, Intel

CSE 571: Probabilistic Robotics 3

Mapping the Allen Center

CSE 571: Probabilistic Robotics 4

C(θ) = dist(i, j)

(i, j)∈ θ

∑

U(θ) = explore(i, j)

(i, j)∈ θ

∑

( )

) ( ) ( max arg θ θ θ

θ

C U − =

∗ [Burgard et al. 00], [Simmons et al. 00]

Coordinated Explora1on

2

Mul1-Robot Mapping With Known Start Loca1ons

CSE 571: Probabilistic Robotics 5

Mul1-Robot Mapping With Known Start Loca1ons

CSE 571: Probabilistic Robotics 6 CSE 571: Probabilistic Robotics 7

Why are Unknown Start Loca1ons Hard?

Robot A Robot B Robot C

► Need to know whether or not maps overlap ► Need to know how maps overlap

Mul1-robot Map Merging

• Problems

– Number of possible merges is exponen1al in number of robots – Cannot merge maps by simply overlaying them

• Wanted

– Scalability, robustness – Merge maps as soon as possible

CSE 571: Probabilistic Robotics 8

3

Mul1-robot Map Merging

CSE 571: Probabilistic Robotics 9

Mul1-robot Map Merging

CSE 571: Probabilistic Robotics 10

Mul1-robot Map Merging

CSE 571: Probabilistic Robotics 11 CSE 571: Probabilistic Robotics 12

Es1ma1ng rela1ve loca1ons

• Idea: Localize one robot in other robot’s map

using par1cle filter

• Problems:

– Only par1al map available – Other robot might be outside the map – Map grows – Impossible to keep track of all loca1ons inside and

• utside the par1al map
• Solu1on: Only keep track of trajectories that
• verlapped map at some 1me

4

CSE 571: Probabilistic Robotics 13

• Overlapping trajectories
• Non-overlapping trajectories

[ ]

∫

− − − − − − − − − − − −

+ ⋅ = ) , | ( ) , | ( ) , | ( ) , | ( ) | ( ) , | (

2 : 1 1 : 1 1 1 1 1 2 : 1 1 : 1 1 1 1 1 : 1 : 1 t t t t t t t t t t t t t t t t t t t

u z n p u n x p dx u z x p u x x p x z p u z x p α

Par1al map localiza1on (intui1on)

) , | ( ) 1 )( | ( ) , | (

2 : 1 1 : 1 1 1 : 1 : 1 − − − −

− =

t t t t t t t t

u z n p

• utside

z p u z n p ε α

CSE 571: Probabilistic Robotics 14

Par1al map localiza1on (example)

CSE 571: Probabilistic Robotics 15

Coordina1on

∑ ∑

∈ ∈

= =

θ θ

θ θ

) , ( ) , (

) , explore( ) ( ) , dist( ) (

j i j i

j i U j i C

∑ ∑

∈ ∈

⎩ ⎨ ⎧ = ⎩ ⎨ ⎧ + =

θ θ

θ θ

) , ( ) , (

hypothesis is if ) , merge( ) ( frontier is if ) , explore( ) ( hypothesis is if ) , meet( ) , dist( frontier is if ) , dist( ) (

j i j i

j j i j p j j i U j j i j i j j i C

( )

) ( ) ( max arg θ θ θ

θ

C U − =

∗ [Burgard et al. 00], [Simmons et al. 00], [Zlot et al. 02]

Hypotheses become potential goals

CSE 571: Probabilistic Robotics 16

Experimental setup

5

CSE 571: Probabilistic Robotics 17

• Rigorously tested by outside evalua1on team
• No tes1ng allowed in 1/2 of environment
• Limited communica1on
• No interven1on / observa1on during experiment
• Comparison to “ground truth” map

CSE 571: Probabilistic Robotics 18

Cen1Bots: Experimental Evalua1on

CSE 571: Probabilistic Robotics 19

Control Center and Test Team

CSE 571: Probabilistic Robotics 20

6

CSE 571: Probabilistic Robotics 21

Comparison to “Ground Truth Map”

CSE 571: Probabilistic Robotics 22

Three Mapping Runs

CSE 571: Probabilistic Robotics 23

Three Overlayed Maps

3D Explora1on

CSE 571: Probabilistic Robotics

[Shen-Michael-Kumar: IIJRR-2012]

Courtesy of Vijay Kumar

24

7

CSE 571: Probabilistic Robotics

¡ EKF with ar1culated ICP over manipulator joint angles, camera

pose and pose of (par1al) object

[Krainin-Henry-Ren-F: IJRR-11]

25

Uncertainty in Object Surface

¡ Signed-distance func1on voxel grid [Curless ’96] ¡ Surface uncertainty from beam-based noise model

CSE 571: Probabilistic Robotics

[Krainin-F-Curless: ICRA-11]

26

View Selec1on Algorithm

• Conceptually similar to Planetarium

Algorithm [Connolly ’85]

• Procedure:

– Generate kinema1cally achievable viewpoints – Compute informa1on gain (quality) for each viewpoint – Select view as tradeoff between quality and cost

CSE 571: Probabilistic Robotics 27

Re-Grasp Selec1on

• Generate candidate grasps [Diankov ‘10]
• Select grasp by maximum informa1on gain,

accoun1ng for occlusion caused by grasp

CSE 571: Probabilistic Robotics 28

8

Mul1ple Grasp Results

• Evaluated regrasping on four
• bjects
• Includes box with three grasps

CSE 571: Probabilistic Robotics 29

Ac1ve Object Modeling

CSE 571: Probabilistic Robotics 30

Ac1ve Mapping

• View selec1on for mapping and segmenta1on

31 CSE 571: Probabilistic Robotics