3 28 3. Towards a task-based computational evaluation benchmark Figure 3.1: The hallway environment in thebaseline study (de Ruyter van Steveninck et al., 2022b) (used with permission). Left: the hallway variation with complex textures. Right: the plain hallway variation. Figure 3.2: The virtual hallway environment. Left: Screenshot of the virtual agent and the environment in the Unity editor. The ceiling is removed for visualization purposes. Center: the hallway with plain textures, seen from the agent perspective. Right: the hallway with complex textures. Environment The environment and task are a virtual reproduction of the (real-world) experiments in the baseline study. The environment is simulated in the Unity game engine (Unity Technologies, CA, USA) and consists of a procedurally generated hallway with boxes that act as obstacles (see Figures 3.1 and3.2). The obstacle placement in the virtual environment is randomized, but constrained to the possible configurations that are found in the baseline study. Textures are created from original photographs and the relative dimensions are kept identical. Due to the virtual nature, there are no limits to the size of the hallway. Therefore, instead of modeling a 22-meter-long mobility course, the hallway is continuously extended and updated with new pieces while moving forward in the environment. The training is broken down into multiple navigation sessions (episodes) that start with a random hallway initialization. To ensure slight variations in the scene representation, the virtual camera is slightly rotated around the forward axis with a random angle on the interval [−3,3] (degrees) at the start of each episode. Task In line with the experiments in the baseline study, the goal for the agent is to move forwards towards the end of the hallway as fast as possible, while avoiding collisions
RkJQdWJsaXNoZXIy MTk4NDMw