7a - Self Localisation

Question 1

What is the key point of self-localisation?

Answer 1

So a robot can find its position and situation in the environment

Question 2

What can be used to approximate functions?

Answer 2

Particle sets

The more particles fall into an interval, the higher the probability of that interval

Question 3

What is a particle?

Answer 3

An individual state estimate

Question 4

What is a particle defined by?

Answer 4

Its state values that determine the robot’s pose in the plane

A weight that indicates its likelihood

Question 5

What do particle filters use?

Answer 5

Many particles to represent the belief state (where am I?)

Question 6

What are the 4 stages to particle filters?

Answer 6

Initialisation, Prediction, Update, Resample

Question 7

What is initialisation in particle filter localisation?

Answer 7

Sampling from initial distribution

No idea where robot is, so throw particles everywhere

Question 8

What are the 3 stages that loop in particle filters?

Answer 8

Prediction, Update, Resample

Question 9

What happens in the prediction step?

Answer 9

For each particle, sample and add random noisy values from the motion model

Question 10

What happens in the update step?

Answer 10

Each particle’s weight is the likelihood of getting the current sensor readings from that particle’s hypothesis (compared to the predicted readings from the map)

Question 11

What happens in the resample step?

Answer 11

A new set of particles is chosen such that each particle survives in proportion to its weight

Survival of the fittest: replace unlikely samples by more likely ones

Question 12

What can particle filters represent?

Answer 12

Arbitrary probability density functions (distributions) using samples

Question 13

What can particle filters solve?

Answer 13

The kidnapped robot problem and handle perceptually aliased environments

Question 14

What is Kalman filter localisation?

Answer 14

Instead of an arbitrary density function, KF uses Gaussians for robot belief, motion and measurement models

Only mean and covariance needs to be updated - efficient computation

Question 15

What does Kalman filter localisation address?

Answer 15

Addresses position tracking, not global localisation or kidnapped robot problem

Question 16

What are the 5 steps to Kalman Filter localisation?

Answer 16

Prediction (ACT) based on previous estimate and odometry

Observation (SEE) with on-board sensors

Measurement prediction based on prediction and map

Matching of observation and map

Estimation -> position update

Question 17

Does does Markov Localisation do?

Answer 17

Uses an explicitly specified probability distribution across all possible robot positions

Question 18

What are the pros of Markov localisation?

Answer 18

Allows for localisation starting from any unknown position

Can recover from ambiguous situations

Question 19

What are the cons of Markov localisation?

Answer 19

Requires a discrete representation of the space, such as a geometric grid of topological graph

Requires memory and computational power

Question 20

What are the pros of Kalman filter localisation?

Answer 20

Tracks the robot from an initially known position

Precise and efficient

Can be used in continuous world representations

Question 21

What are cons to Kalman filter localisation?

Answer 21

If uncertainty of robot becomes large (collision with object), it can fail to capture the multitude of possible robot positions and become irrevocably lost