Behavior: Robot manipulation benchmark based on 1000 household tasks (github.com)

Stanford’s BEHAVIOR-1K is a new, large-scale simulation benchmark and monolithic repo for embodied AI: it packages everything needed to train and evaluate agents on 1,000 everyday household activities (cleaning, cooking, organizing) curated from human time‑use surveys and preference studies. The release is meant as a human‑centered evaluation suite that stresses long-horizon, multi-step, object-centric tasks and supports teleoperation, evaluation tooling and action primitives—making it a practical standard for research on generalization, planning, imitation learning and RL in realistic home settings. Technically, BEHAVIOR-1K provides a modular installation script (recommended stable tag v3.7.1) that installs core components such as OmniGibson (physics & sim), BDDL (Behavior Domain Definition Language for task specs) and JoyLo (teleop interface). Target platforms are Linux (Ubuntu 20.04+) and Windows 10+; recommended hardware is 32+ GB RAM, 8+ GB VRAM and an NVIDIA RTX 2080+ GPU. Flags let you choose new conda envs, datasets, primitives, eval/dev deps and CUDA version (default 12.4); there are non‑interactive options to accept Conda/NVIDIA/dataset EULAs for CI. By combining scale, realistic simulation and standardized evaluation, BEHAVIOR-1K aims to improve reproducibility and push embodied agents toward robust, human‑relevant everyday competence.