In the LightDark2D problem, cost is quadratic in the distance from the origin; In the LightDark2DTarget problem, there is a cost of -1 accrued for every step outside of the target region (a radius from the origin controlled by the term_radius field of LightDark2DTarget).
There is also a Kalman filter LightDark2DKalman, but Kalman filtering is rather ill-suited to this problem because the true beliefs are not similar to a Gaussian distribution. Instead, it is much better to use a particle filter from ParticleFilters.jl.
See test/runtests.jl for usage examples.