NVlabs releases SAGE for embodied AI training
NVlabs published SAGE, a tool that generates richly detailed 3D scenes intended for embodied AI training and simulation data generation. The release is meant to accelerate sim‑to‑real by producing diverse, physically plausible environments for agents to learn in. (x.com)
NVlabs published the SAGE‑10k dataset containing 10,000 interactive indoor scenes spanning 50 room types and roughly 565,000 uniquely generated 3D objects. (nvlabs.github.io) The SAGE paper (arXiv:2602.10116) lists authors Hongchi Xia, Xuan Li, Zhaoshuo Li, Qianli Ma, Jiashu Xu, Ming‑Yu Liu, Yin Cui, Tsung‑Yi Lin, Wei‑Chiu Ma, Shenlong Wang, Shuran Song, and Fangyin Wei, with the first submission on Feb 10, 2026 and a revision on Feb 20, 2026. (arxiv.org) The repository includes an explicit NVIDIA Isaac Sim integration under server/isaacsim and components labeled M2T2 and robomimic for generating contact‑rich manipulation data and training policies. (github.com) Files in the repo show connectors to foundation models and asset synthesis tools—examples include a foundation_models.py module, a matfuse‑sd directory for material generation, and TRELLIS integrations for 3D asset creation. (github.com) The paper reports that policies trained purely on SAGE‑10k exhibit measurable scaling trends and generalize to unseen objects and layouts in the authors’ experiments. (arxiv.org) NVIDIA lists SAGE as a CVPR 2026 project on its Deep Imagination Research page, and the public code release is provided under an Apache‑2.0 license in the NVlabs GitHub repository. ( ) The project page and repo package both scene and action generation code and provide a direct download link for the SAGE‑10k dataset to enable immediate use in simulator workflows. ( )
