Habitat-Sim is a high-performance 3D simulator for embodied AI research.
The simulator addresses the need for fast, physics-enabled environments where embodied agents can learn and be evaluated. It prioritizes simulation speed over breadth of capabilities, achieving several thousand frames per second when rendering scenes from standard datasets and over 8,000 steps per second when simulating robot interactions with rigid-body dynamics. The tool supports 3D scans of indoor and outdoor spaces from datasets like HM3D, MatterPort3D, Gibson, and Replica, as well as CAD models of spaces and objects. It includes configurable sensors such as RGB-D cameras and egomotion sensing, can simulate robots described via URDF including mobile manipulators, fixed-base arms, and quadrupeds, and uses Bullet for rigid-body mechanics.
Habitat-Sim is designed for researchers building embodied AI systems who need fast simulation for training and evaluation at scale. It suits projects requiring physics simulation of robot interactions in realistic 3D environments, particularly when training agents through reinforcement learning or imitation learning. The tool is typically used alongside Habitat-Lab, a companion library that provides high-level task definitions, agent training pipelines, and benchmarking metrics for embodied AI experiments. Researchers working with navigation, manipulation, or instruction-following tasks in photorealistic or CAD-based environments will find the performance characteristics most valuable.
The project is no longer receiving active development or maintenance by the original team, though the codebase remains available for independent use and modification. The simulator has been iterated through multiple major versions with corresponding research publications. External contributions are documented and credited within the project. The tool maintains integration with established datasets and robot platforms, supporting reproducible research through standard citation practices.
