Habitat: A Platform for Embodied AI Research
Manolis Savva1,4*, Abhishek Kadian1*, Oleksandr Maksymets1*, Yili Zhao1, Erik Wijmans1,2,3, Bhavana Jain1, Julian Straub2, Jia Liu1, Vladlen Koltun5, Jitendra Malik1,6, Devi Parikh1,3, Dhruv Batra1,3 1Facebook AI Research, 2Facebook Reality Labs, 3Georgia Institute of Technology, 4Simon Fraser University, 5Intel Labs, 6UC Berkeley https://aihabitat.org
Abstract
We present Habitat, a platform for research in embodied artificial intelligence (AI). Habitat enables training embodied agents (virtual robots) in highly efficient photorealistic 3D simulation. Specifically, Habitat consists of: (i) Habitat-Sim: a flexible, high-performance 3D simulator with configurable agents, sensors, and generic 3D dataset handling. Habitat-Sim is fast – when rendering a scene from Matterport3D, it achieves several thousand frames per second (fps) running single-threaded, and can reach over $10@000$ fps multi-process on a single GPU. (ii) Habitat-API: a modular high-level library for end-to-end development of embodied AI algorithms – defining tasks (e.g. navigation, instruction following, question answering), configuring, training, and benchmarking embodied agents.
中文速览
虚拟机器人要在真实室内场景里学会自主导航，但现有仿真平台速度太慢、模块耦合太紧，导致大规模训练实验根本无法开展。为此，研究者推出了 Habitat 平台，核心包括高性能 3D 仿真器 Habitat-Sim（单线程即可达数千帧/秒，多进程在单张 GPU 上超过 10,000 帧/秒，比同类平台快 2–3 个数量级）和模块化开发库 Habitat-API（支持任务定义、智能体训练与基准测试全流程）。凭借这套平台，研究者首次将强化学习智能体的训练经验扩展到 7500 万步，发现当训练量达到足够规模时，学习方法的导航性能可以超越传统 SLAM 方法，推翻了此前规模较小实验得出的相反结论；同时开展了首批跨数据集泛化实验，发现只有配备深度（D）传感器的智能体才能在 Matterport3D 与 Gibson 两个数据集之间稳定泛化。这一开源平台为具身 AI 研究提供了统一、高效、可复现的基础设施，有望成为推动该领域进展的共同基准。
原文 arXiv:1904.01201;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1904.01201v2