opof-grid2d

OPOF simple navigation domains in a 2D grid world to help users familiarize with OPOF. They also act as a sanity check for developing optimization algorithms.

opof-grid2d is maintained by the Kavraki Lab at Rice University.

Installation

$ pip install opof-grid2d

opof-grid2d is officially tested and supported for Python 3.9, 3.10, 3.11.

Domain: RandomWalk2D[size]

from opof_grid2d.domains import RandomWalk2D
domain = RandomWalk2D(11) # Creates a RandomWalk2D domain instance for a 11x11 board.

Description

An agent starts at a location (green) and moves in random directions according to some fixed probabilities until it reaches the goal (magenta). When attempting to move into an obstacle (black) or the borders of the grid, a step is spent but the position of the agent does not change. The probability of moving in each direction is fixed across all steps.

Planner optimization problem

We want to find a generator that maps a problem instance $c$ (in this case, the combination of board layout and start and goal positions) into direction probabilities (in this case, vectors $\in \mathbb{R}^4$ with non-negative entries summing to $1$), such that the number of steps taken during the random walk is minimized.

Planning objective

$\boldsymbol{f}(x; c)$ is given as $- steps / (4 \times size^2)$, where $steps$ is the number of steps taken to reach the goal. A maximum of $4 \times size^2$ steps are allowed.

Problem instance distribution

The training set and testing set each contain $1000$ problem instances, where the obstacle, start, and goal positions are uniformly sampled.

Domain: Maze2D[size]

from opof_grid2d.domains import Maze2D
domain = Maze2D(11) # Creates a Maze2D domain instance for a 11x11 board.

Description

A* search is run against a heuristic function $h(n)$ to find a path from the start (green) to the goal (green). The heuristic function determines the priority in which nodes are expanded (cells that are in darker red have a lower $g(n) + h(n)$ value, and have higher priority). The maze is assumed to be perfect, i.e., there is exactly one path between any two cells.

Planner optimization problem

We want to find a generator $G_\theta(c)$ that maps a problem instance $c$ (in this case, the combination of board layout and start and goal positions) to $h(n)$ (in this case, assignments of values $\in [0, size^2]$ to each of the $size^2$ cells) such that the number of nodes expanded in the A* search is minimized.

Planning objective

The planning objective $\boldsymbol{f}(x; c)$ is given as $- steps / n_{\mathrm{empty}}$, where $steps$ is the number of nodes expanded before finding the goal and $n_{\mathrm{empty}}$ is the number of obstacle-free cells.

Problem instance distribution

The training set and testing set each contain $1000$ problem instances, where the maze is generated using Wilson's algorithm, and the start and goal positions are uniformly sampled.

Citing

TBD

License

opof-grid2d is licensed under the BSD-3 license.

opof-grid2d is maintained by the Kavraki Lab at Rice University, funded in part by NSF RI 2008720 and Rice University funds.