readme

SLAM Gridsim use with sparer sensor

**********************************************************************
* gridsim2 -- occupancy mapping simulation with a generic sensor model
* Kris Beevers
* [e-mail]
* 11/08/2006
**********************************************************************

This code implements a simulation for the environment, robot, and
sensor models described in the following publications:

K. Beevers and W. Huang.  How sensing capabilities influence map
quality.  Submitted to ICRA 2007.

K. Beevers.  Mapping with limited sensing.  PhD thesis, Rensselaer
Polytechnic Institute, January 2007.

The main purpose of the simulations is to validate our analytical
bounds.

Too see some of the parameters you can set, run ./sim --help after
compiling.  (To compile, just run make.)  Most of the parameters are
for the sensor model.  You can also specify an existing environment
file (pgm format), or set some parameters for how to generate the
environment, e.g., independent cell occupancy probabilities, or MRF
model with Gibbs sampling.  There are some example MRF models (*.mrf).
You can also specify the type of trajectory, either is a pre-generated
pose sequence, a waypoint file to be fed to a simple controller, or
random generation of either a realistic trajectory (based on the
simple controller) or just a set of random poses.  Our analytical
model uses random poses but the simulations show that there is
generally not much difference in map error between random poses and a
more realistic trajectory.

There are a variety of configuration files with different sensor model
parameters.  These correspond to the synthetic sensors described in
the paper.

Unless the --nofiles option is specified, the simulation outputs
several images in pgm format:

  * grid.pgm: the ground truth map
  * map.pgm: the occupancy map produced by the sensor model
  * map-ml.pgm: the maximum likelihood map
  * possible.pgm: the best possible map (recognizes occluded cells)
  * real-freq.pgm: the true frequency of observations of each cell
  * update-freq.pgm: the frequency of updates of each cell

The simulation also prints out several numbers.  In order these are:

  * the actual ground truth grid density
  * the entropy of the ground truth grid with respect to 4-neighbors
  * squared error of map.pgm wrt grid.pgm, i.e. sum(abs(map[i]-grid[i])^2)
  * squared error of map-ml.pgm wrt grid.pgm
  * mean num. real observations of each cell
  * mean num. updates of each cell
  * mean ratio of real observations to updates

Note that in computing the squared error of the maximum likelihood
map, a fair coin is flipped for cells with likelihood ~1/2 to
determine the ML occupancy.

As the code is fairly short and simple, the best way to get an idea of
what is going on is to simply browse it.

There are also a few simple utility programs:

  * diff: compute the difference between two pgm grids
  * gridstats: compute density and 4-entropy of a pgm grid
  * maxmap: compute the ML map from an occupancy grid
  * montecarlo.pl: run a bunch of simulations (see the code)
  * model_error3.m: matlab implementation of the bounds from the paper
  * makefigs2.m: makes some plots from simulated data
  * real_sensors.m: applies analytical model to realistic sensors
  * makerobot.m: makes a picture of the sensor for given parameters

(Other matlab scripts are mainly supporting code.)