Training robots and visual perception models in the physical world is slow, expensive, and risky. Simulation offers a faster alternative, but models trained in synthetic environments often fail to transfer to real conditions. Gibson Env is a virtual environment designed to close that gap by grounding simulation in reality.
Gibson is built by virtualizing actual buildings rather than designing artificial scenes. It includes more than 1,400 floor spaces drawn from 572 full buildings, giving agents access to the complexity of real-world layouts. Agents are also embodied, meaning they must respect physical constraints such as space and movement.
A key component is 'Goggles,' an internal synthesis mechanism that lets models trained in Gibson be deployed in the real world without additional domain adaptation. The paper, presented at CVPR 2018 and later revised, demonstrates sample perceptual tasks learned in the environment. With a single source, there are no conflicting accounts to compare; the summary above reflects the paper's own claims.