A world is everything around the drone — walls, floor, obstacles, lighting, gravity. The drone model itself is loaded into a world via <include>. Different worlds are used for different purposes: smoke-testing, RL training, sim-to-real evaluation.
Three pages here. Empty world is the minimal world for hovering tests — flat floor, no obstacles, default lighting. Useful for “does the drone fly at all” smoke testing and for the baseline reference in benchmarks. Obstacle maze is the more interesting training world — corridors, pillars, the kind of geometry the RL policy actually needs to learn to navigate. Dynamic wind world adds time-varying wind for domain randomization — testing whether the policy generalizes to disturbances it didn’t see during training.
The flagship training world for the current RL experiments is warehouse_v2, which is part of the obstacle-maze family but with more realistic indoor warehouse geometry. Its specific details live in the RL framework section rather than here, since the geometry was chosen to match the last-mile delivery and SAR use cases.
Contents
Auto-generated from child entries during build (update-indexes.mjs).