SLAM

Question 1

When is SLAM used?

Answer 1

• need autonomy
• no prior map
• don’t want to use GPS

Question 2

SLAM assumptions

Answer 2

• map and localise at same time so assume only robot moves and world is static

Question 3

Algorithm for SLAM

Answer 3

1. robot starts has 0 uncertainty
2. robot drives and uncertainty grows
3. robot initialises obstacles
4. obstacles inherit uncertainty of robot + extra
5. robot goes back to start uncertainty grows
6. robot remeasures start position, uncertainty decreases
7. robot remeasures one obstacle, uncertainty of both decrease

Question 4

Limits of SLAM

Answer 4

• poor computational scaling
• growth in uncertainty
• data association hard at high uncertainty

Question 5

Large scale SLAM algorithm

Answer 5

1. local metric mapping - estimate trajectory and make local map
2. place recognition - loop closure
3. map optimisation - optimise closed loops

Question 6

How does loop closure detection work

Answer 6

save images at regular intervals and use image retrieval

Question 7

How does pure topological SLAM work?

Answer 7

keep record of places visited and how they connect without explicit geometry info

Question 8

What is pose graph optimisation?

Answer 8

computes set of node positions which is maximally probable given metric and topological constraints

Question 9

What are factor graphs?

Answer 9

• each factor is likelihood of one measurement (dot)
• factors depend on subset of variables (circles) in graph
• total likelihood of all measurements is product of all factors

Question 10

How does factor graph inference work?

Answer 10

find most probable variable values given measurements

Question 11

What are methods of factor graph inference?

Answer 11

global batch optimisation: gradient of likelihood w.r.t. variables and adjust until maximum
incremental filtering
incremental peice-wise optimisation
distributed inference via Gaussian Belief Propagation