GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

NVIDIA, :, Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi "Jim" Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, Joel Jang, Zhenyu Jiang, Jan Kautz, Kaushil Kundalia, Lawrence Lao, Zhiqi Li, Zongyu Lin, Kevin Lin, Guilin Liu, Edith Llontop, Loic Magne, Ajay Mandlekar, Avnish Narayan, Soroush Nasiriany, Scott Reed, You Liang Tan, Guanzhi Wang, Zu Wang, Jing Wang, Qi Wang, Jiannan Xiang, Yuqi Xie, Yinzhen Xu, Zhenjia Xu, Seonghyeon Ye, Zhiding Yu, Ao Zhang, Hao Zhang, Yizhou Zhao, Ruijie Zheng, Yuke Zhu · N/A · 2025

Summary

Abstract Summary

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.

Key Points

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

GR00T N1 Foundation Model

GR00T N1 is a Vision-Language-Action (VLA) model for humanoid robots trained on diverse data sources. The model contains a vision-language backbone that encodes language and image input and a DiT-based flow-matching policy that outputs high-frequency actions. We use the NVIDIA Eagle-2 VLM ( Li et al., 2025 ) as the vision-language backbone.

Specifically, our publicly released GR00T-N1-2B model has 2.2B parameters in total, with 1.34B in the VLM. The inference time for sampling a chunk of 16 actions is 63.9ms on an L40 GPU using bf16. Fig.

GR00T N1 Foundation Model — GR00T N1: An Open Foundation Model for Generalist Humanoid Robots — Figure 2 : GR00T N1 Model Overview. Our model is a Vision-Language-Action (VLA) model that adopts a dual-system design. We convert the image observation and language instruction into a sequence of tokens to be processed by the Vision-Language Model (VLM) backbone. The VLM outputs, together with robot state and action encodings, are passed to the Diffusion Transformer module to generate motor actions.

Pre-Training Datasets

We structure our pre-training corpus into three main categories: real-robot datasets (Sec. 3.1 ), synthetic datasets (Sec. 3.2 ), and human video datasets (Sec. 3.3 ). These roughly correspond to the peak, middle, and base of the data pyramid (Fig. 1 ), respectively. The synthetic datasets consist of both simulation trajectories and neural trajectories.

Table 1 summarizes our training data generation strategies in Sec. 2.2 and their applicable data sources correspondingly.

Pre-Training Datasets — GR00T N1: An Open Foundation Model for Generalist Humanoid Robots — Figure 3 : GR00T N1 Model Architecture. GR00T N1 is trained on a diverse set of embodiments ranging from single-arm robot arms to bimanual humanoid dexterous hands. To deal with different robot embodiment’s state observation and action, we use DiT blocks with an embodiment-aware state and action encoder to embed the robot’s state and action inputs. GR00T N1 model leverages latent embeddings of the Eagle-2 model to incorporate the robot’s visual observation and language instructions. The vision language tokens will then be fed into the DiT blocks through cross-attention layers.

Evaluation

We evaluate our GR00T N1 models in a diverse set of simulated and real-world benchmarks. Our simulation experiments are conducted on three distinct benchmarks designed to systematically assess the effectiveness of our model across various robot embodiments and manipulation tasks. In our real-world experiments, we investigate the model’s capability on a suite of tabletop manipulation tasks with the GR-1 humanoid robot.

These experiments aim to demonstrate GR00T N1’s ability to acquire new skills from a limited number of human demonstrations.

Evaluation — GR00T N1: An Open Foundation Model for Generalist Humanoid Robots — Figure 5 : Synthetically Generated Videos. We leverage off-the-shelf video generation models to create neural trajectories to increase the quantity and diversity of our training datasets. These generated data can be used for both pre- and post-training of our GR00T N1. (1) The first three rows are generated from the same initial frames but with different prompts (change left or right, the location to place the object), (2) the following two are from the same initial frames but replace the object to pick up, (3) the next row showcases the video model generating a robot trajectory which is very challenging to generate in simulation (spilling contents inside a mesh cup into a bin), and (4) the last row is generated from an initial frame from simulation data. We use the red rectangles to indicate the initial frames.

Sources and Demonstrations

Figures are reproduced from NVIDIA et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.