habitat-lab

Why we included this project

Habitat-Lab pairs the Habitat-Sim renderer with a higher-level layer for defining tasks, equipping embodied agents with sensors, and running single- or multi-agent training through reinforcement or imitation learning. That arrangement lets researchers move from scene and agent definitions to trained policies without bolting everything together by hand, and it covers the usual indoor embodied tasks, from navigation and rearrangement to instruction following and human-robot interaction. The library also ships ready baselines and standard evaluation metrics, so benchmarking a new idea is more direct than building the harness from scratch. One practical caveat: Meta has stopped actively maintaining the project past v0.3.4, so teams should plan to fork and maintain their own copy. Even so, it remains one of the most complete references for building embodied-agent experiments on photorealistic indoor scenes.