Flatland

Flatland provides a framework which helps to investigate and test various new approaches to the VRSP. Instead of a complete physics based simulation, Flatland reduces the complexity by simulating agents in a 2D-Grid environment. It provides ways to generate environments, handle agent movements based on action spaces and observations, simulate disruptive events, implement machine learning techniques and more.

Agents

In Flatland trains are represented as agents with a specific position and orientation. These agents are able to traverse the environment by selecting a single action from their action space at each time step. This action space depends on the transition map of the cell and the occupation of neighbouring cells by other agents.

Each agent has the goal to travel, optimally with minimal delay, from their starting location to their target location which both are defined by train stations in cities.

Grid

The grid describes the world in Flatland. It is a matrix of cells of a specific type and orientation which, using transition maps, define if and how an agent can traverse them.

Transition Maps

Each cell's possible transitions are defined by four 4-bit transition maps. Each map represents an orientation of an agent in a specific direction (north, east, south, west) while each bit describes if traversal in one direction is possible or not.

For example the transition maps

1000 0000 0010 0000

describe a cell in which a north-oriented agent can go north, a west or east oriented agent can not traverse at all and a south oriented agent can traverse south.

In other words this is a straight line going from south to north.

Cell types

An empty cell is the most basic cell, its transition maps allow no transition and it can't be occupied by an agent at any time. All other cells represent tracks, each track cell can be occupied by at most one agent at a time.

The different track types are:

• straight lines (south-north)
• simple switches (straight south-north and a south to west turn)
• diamond crossing (two straight lines crossing, no turns possible)
• single slip switch (simple switch and diamond crossing)
• double slip switch (diamond crossing, simple switch and a simple switch rotated by 180 degrees)
• symmetrical switch (south-to-west turn, south-to-east turn)
• dead end (no further traversal possible, agent has to turn around)

These cell types are sufficient to build railway networks of complexity comparable to real-world examples by either using the base version as mentioned or use rotated variants of these base versions.