GibsonEnv

Why we included this project

Gibson is a simulator for embodied AI research that builds its environments from real scanned indoor spaces instead of hand-modeled synthetic ones, so agents see the same clutter and lighting irregularities they would encounter in actual buildings. It couples perception with movement through a Bullet physics engine that enforces collision and body constraints, and includes a 'Goggles' mode designed to help policies learned in simulation transfer to a physical robot. The repository carries setup instructions, Docker support, ROS integration for real robot workflows, and a database of hundreds of spaces, with a 39-space subset bundled so you can run initial experiments without downloading everything. For teams training navigation or manipulation policies and aiming for sim-to-real transfer, Gibson remains a relevant reference point, and the CVPR 2018 paper gives a solid explanation of the design choices.