General-purpose robots need a versatile body and an intelligent mind. Recent advancements in humanoid robots have shown great promise as a hardware platform for building generalist autonomy in the human world. A robot foundation model, trained on massive and diverse data sources, is essential for enabling the robots to reason about novel situations, robustly handle real-world variability, and rapidly learn new tasks. To this end, we introduce GR00T N1, an open foundation model for humanoid robots. GR00T N1 is a Vision-Language-Action (VLA) model with a dual-system architecture. The vision-language module (System 2) interprets the environment through vision and language instructions. The subsequent diffusion transformer module (System 1) generates fluid motor actions in real time. Both modules are tightly coupled and jointly trained end-to-end. We train GR00T N1 with a heterogeneous mixture of real-robot trajectories, human videos, and synthetically generated datasets. We show that our generalist robot model GR00T N1 outperforms the state-of-the-art imitation learning baselines on standard simulation benchmarks across multiple robot embodiments. Furthermore, we deploy our model on the Fourier GR-1 humanoid robot for language-conditioned bimanual manipulation tasks, achieving strong performance with high data efficiency.
核心贡献 · Key contributions
提出 GR00T N1，一个面向人形机器人的开放基础 VLA 模型，采用双系统架构。 Introduces GR00T N1, an open foundation VLA model for humanoid robots with dual-system architecture.
开发了数据金字塔协同训练策略，融合真实机器人、人类视频和合成数据。 Develops a data pyramid co-training strategy using real robot, human video, and synthetic data.
在多个机器人形态的仿真基准上实现了最先进的模仿学习结果。 Achieves state-of-the-art imitation learning results on simulation benchmarks across multiple embodiments.
在 Fourier GR-1 人形机器人上展示了高数据效率的强真实世界性能。 Demonstrates strong real-world performance on Fourier GR-1 humanoid with high data efficiency.
为社区提供开源模型权重、训练数据和仿真基准。 Provides open-source model checkpoint, training data, and simulation benchmarks for community.
局限 · Limitations
主要关注短时域桌面操作任务，而非长时域移动操作。 Focuses primarily on short-horizon tabletop manipulation tasks, not long-horizon loco-manipulation.
合成数据生成在物理一致性和多样性方面仍面临挑战。 Synthetic data generation still faces challenges in physics adherence and diversity.
当微调数据分布与预训练显著偏离时，模型性能可能下降。 Model performance may degrade when fine-tuning data distribution shifts significantly from pre-training.
真实世界评估限于特定任务和机器人硬件；对其他领域的泛化性未验证。 Real-world evaluation limited to specific tasks and robot hardware; generalization to other domains not verified.
论文章节 · Sections（共 28）
摘要Abstract
1 引言1 Introduction
2 GR00T N1 基础模型2 GR00T N1 Foundation Model
2.1 模型架构2.1 Model Architecture
2.2 训练数据生成2.2 Training Data Generation
2.3 训练细节2.3 Training Details
3 预训练数据集3 Pre-Training Datasets
3.1 真实世界数据集3.1 Real-World Datasets
3.2 合成数据集3.2 Synthetic Datasets
3.3 人类视频数据集3.3 Human Video Datasets
4 评估4 Evaluation
4.1 仿真基准4.1 Simulation Benchmarks
4.2 真实世界基准4.2 Real-World Benchmarks
4.3 实验设置4.3 Experiment Setup
4.4 定量结果4.4 Quantitative Results
4.5 定性结果4.5 Qualitative Results
4.6 局限性4.6 Limitations
5 相关工作5 Related Work
6 结论6 Conclusions
A.1 核心贡献者A.1 Core Contributors
A.2 贡献者A.2 Contributors
A.3 致谢A.3 Acknowledgments
附录 B 详细实验结果Appendix B Detailed Experiment Results
附录 C 额外定性结果Appendix C Additional Qualitative Results