Jethro's Braindump

Simultaneous Localization and Mapping

In SLAM, the robot acquires a map of its environment while simultaneously localizing itself relative to this map. It is a significantly more difficult problem compared to Robot Localization and Occupancy Grid Mapping.

There are 2 main forms of SLAM:

Online SLAM
estimating the posterior over the momentary pose along with the map: \(p(x_t, m | z_{1:t}, u_{1:t})\)
full SLAM
posterior over the entire path \(x_{1:t}\) along with the map: \(p(x_{1:t}, m | z_{1:t}, u_{1:t})\)

The online SLAM algorithm is the result of integrating out past poses in the full SLAM problem

EKF SLAM

EKF SLAM uses a number of approximations and limiting assumptions:

Feature-based maps
Maps are composed of a small number of point landmarks. The method works well when the landmarks are relatively unambiguous.
Gaussian Noise
The noise in motion and perception is assumed to be Gaussian.
Positive Measurements
It can only process positive sightings of landmarks, ignoring negative information.

SLAM with Known Correspondence

The key idea is to integrate landmark coordinates into the state vector. This corresponds to the continuous portion of the SLAM problem. EKF SLAM combines the state vector as such:

\begin{aligned} y_{t} &=\left(\begin{array}{c}{x_{t}} \ {m}\end{array}\right) \end{aligned}

and calculates the online posterior:

\begin{equation} p\left(y_{t} | z_{1: t}, u_{1: t}\right) \end{equation}

EKF SLAM with Unknown Correspondences

It uses an incremental maximum likelihood estimator to determine correspondences.

Feature Selection and Map Management

EKF SLAM requires several additional techniques to be robust in practice. First, one needs to deal with outliers in the measurement space. One technique is to maintain a provisional landmark list. Instead of adding the landmark immediately, it is added to this list, and when the uncertainty has shrunk after repeated observations of the landmark, it is added in.

EIF SLAM

Unlike EKF SLAM, the extended information form SLAM algorithm (EIF SLAM) solves the full SLAM problem. EIF represents the posterior gaussian in its canonical representation form, with the precision matrix and information state vector (See Information Filter).

EIF SLAM is also not incremental: it calculates posteriors over a robot path. It is best suited for problems where a map needs to be built from data of fixed size, and can afford to hold the data in memory until the map is built.

Suppose we are given a set of measurements \(z_{1:t}\) with associated correspondence variables \(c_{1:t}\), and a set of controls \(u_{1:t}\). Then the EIF SLAM algorithm operates as follows:

  1. Construct the information matrix and information vector fromt he joint space of robot poses \(x_{1:t}\) and map \(m = \{m_j\}\).
  2. Each measurement leads to a local update of \(\Omega\) and \(\xi\). This is because information is an additive quantity.

The key insight is that information is sparse. Specifically,

  • Measurements provide information of a feature relative to the robot’s pose at the time of measurement, forming constraints between pairs of variables.
  • Motion provides information between two subsequent poses, also forming constraints

EIF SLAM records all this information, through links that are defined between poses and features, and pairs of subsequent poses. However, this information representation does not provide estimates of the map or robot path.

Maps are recovered via an iterative procedure involving 3 steps:

  1. Construction of a linear information form through Taylor expansion
  2. Reduction of this linear information form
  3. Solving the resulting optimization problem

Sparse EIF SLAM

The sparse EIF SLAM only maintains a posterior over the present robot pose and the map. Hence, they can be both run online, and are efficient. Unlike EKFs, they also maintain information representation of all knowledge.