habitat-sim

Why we included this project

Habitat-Sim has long been the workhorse of embodied AI research, and it earns that reputation by being fast. It loads real-world scans and CAD scenes into physics-enabled environments, with pluggable RGB-D sensors, egomotion sensing, and URDF-defined robots such as mobile manipulators and quadrupeds, so you can set an agent loose and watch what your policy actually does. That speed matters when training reinforcement learning agents demands many thousands of steps per second. Habitat-Sim is not a turnkey application, though: it works with the companion Habitat-Lab library for building full experiments, so expect to wire the two together. One caveat worth planning around is that Meta stopped active maintenance past a certain version, so budget for a pinned or community-forked setup rather than regular upstream fixes.