Habitat is a platform for research in embodied artificial intelligence (AI), enabling the training of embodied agents (virtual robots) in highly efficient photorealistic 3D simulations. The platform consists of Habitat-Sim, a flexible, high-performance 3D simulator with configurable agents, sensors, and generic 3D dataset handling, and Habitat-API, a modular high-level library for end-to-end development of embodied AI algorithms. Habitat-Sim can render scenes at thousands of frames per second, significantly faster than previous simulators, and supports a wide range of 3D datasets. Habitat-API allows the definition and evaluation of a broad set of tasks, including navigation, instruction following, and question answering. The platform enables large-scale engineering contributions that allow answering scientific questions requiring experiments that were previously impractical. In the context of point-goal navigation, the study revisits the comparison between learning and SLAM approaches and finds that learning outperforms SLAM when scaled to a larger amount of experience. Additionally, the study conducts the first cross-dataset generalization experiments, finding that only agents with depth sensors generalize well across datasets. Habitat aims to support a complementary research program by standardizing the entire 'software stack' for training embodied agents, including generic dataset support, a highly performant simulator, and a flexible API. The platform addresses the challenges of training robots in the real world, such as being slow, dangerous, resource-intensive, and difficult to control. Habitat enables training in simulations, which are faster, safer, and more cost-effective, and allows for fair comparison and benchmarking of progress in a community-wide effort. The Habitat platform has been used to demonstrate the utility of the design in experiments testing the generalization of goal-directed visual navigation agents between different environments and comparing the performance of learning-based agents against classic agents. The study found that learning-based agents significantly outperform SLAM agents, and that agents with depth sensors generalize better across datasets. The results highlight the importance of depth sensors in navigation tasks and suggest that visual navigation agents could benefit from curriculum learning. The Habitat Challenge is an autonomous navigation challenge aimed at benchmarking and advancing efforts in goal-directed visual navigation. The challenge requires participants to upload code rather than predictions, enabling evaluation in novel test environments. The challenge infrastructure leverages the EvalAI platform and allows participants to train their agents in any language, framework, or infrastructure. The challenge is designed to reduce the gap between simulation and reality and increase the difficulty of the task. Future work includes incorporating physics simulation, enabling physics-based interaction between agents and objects, and focusing on distributed simulation settings with large numbers of agents. The platform aims to unify existing community efforts and accelerate research into embodied AI.Habitat is a platform for research in embodied artificial intelligence (AI), enabling the training of embodied agents (virtual robots) in highly efficient photorealistic 3D simulations. The platform consists of Habitat-Sim, a flexible, high-performance 3D simulator with configurable agents, sensors, and generic 3D dataset handling, and Habitat-API, a modular high-level library for end-to-end development of embodied AI algorithms. Habitat-Sim can render scenes at thousands of frames per second, significantly faster than previous simulators, and supports a wide range of 3D datasets. Habitat-API allows the definition and evaluation of a broad set of tasks, including navigation, instruction following, and question answering. The platform enables large-scale engineering contributions that allow answering scientific questions requiring experiments that were previously impractical. In the context of point-goal navigation, the study revisits the comparison between learning and SLAM approaches and finds that learning outperforms SLAM when scaled to a larger amount of experience. Additionally, the study conducts the first cross-dataset generalization experiments, finding that only agents with depth sensors generalize well across datasets. Habitat aims to support a complementary research program by standardizing the entire 'software stack' for training embodied agents, including generic dataset support, a highly performant simulator, and a flexible API. The platform addresses the challenges of training robots in the real world, such as being slow, dangerous, resource-intensive, and difficult to control. Habitat enables training in simulations, which are faster, safer, and more cost-effective, and allows for fair comparison and benchmarking of progress in a community-wide effort. The Habitat platform has been used to demonstrate the utility of the design in experiments testing the generalization of goal-directed visual navigation agents between different environments and comparing the performance of learning-based agents against classic agents. The study found that learning-based agents significantly outperform SLAM agents, and that agents with depth sensors generalize better across datasets. The results highlight the importance of depth sensors in navigation tasks and suggest that visual navigation agents could benefit from curriculum learning. The Habitat Challenge is an autonomous navigation challenge aimed at benchmarking and advancing efforts in goal-directed visual navigation. The challenge requires participants to upload code rather than predictions, enabling evaluation in novel test environments. The challenge infrastructure leverages the EvalAI platform and allows participants to train their agents in any language, framework, or infrastructure. The challenge is designed to reduce the gap between simulation and reality and increase the difficulty of the task. Future work includes incorporating physics simulation, enabling physics-based interaction between agents and objects, and focusing on distributed simulation settings with large numbers of agents. The platform aims to unify existing community efforts and accelerate research into embodied AI.