pomdp_py.problems.light_dark package


A simple continuous domain.

Reference: Belief space planning assuming maximum likelihood observations.

Quoting from the original paper on problem description:

In the light-dark domain, a robot must localize its position in the plane before approaching the goal. The robot’s ability to localize itself depends upon the amount of light present at its actual position. Light varies as a quadratic function of the horizontal coordinate. Depending upon the goal position, the initial robot position, and the configuration of the light, the robot may need to move away from its ultimate goal in order to localize itself. Figure 1 illustrates the configuration of the light-dark domain used in our experiments. The goal position is at the origin, marked by an X in the figure. The intensity in the figure illustrates the magnitude of the light over the plane. The robot’s initial position is unknown.

Figure from the paper

Light-Dark domain

Note that a variant of this domain appears in Online algorithms for POMDPs with continuous state, action, and observation spaces where the state and action spaces are discrete and the observation space is kept continuous. Here we do not implement this variant; You can implement it and create a pull request!

There is an implementation of a variant of LightDarkDomain in the JuliaPOMDP suite (repo).

BLQR planning implementation not complete yet.



pomdp_py.problems.light_dark.problem_discrete module

pomdp_py.problems.light_dark.test module

Module contents