GROOT N1: An Open Foundation Model for Generalist Humanoid Robots (March 2025)

Title: GR00T N1: An Open Foundation Model for Generalist Humanoid Robots (Mar 2025)
Link: arxiv.org/abs/2503.14734v2
Date: March 2025

Summary:
The paper introduces GROOT N1, an open foundation model for humanoid robots. It's a Vision-Language-Action (VLA) model with a dual-system architecture trained on a heterogeneous mixture of real-robot trajectories, human videos, and synthetic datasets. The model outperforms state-of-the-art imitation learning baselines and demonstrates strong performance in real-world manipulation tasks on the Fourier GR-1 humanoid robot.

Key Topics:
Foundation Models
Humanoid Robots
Robotics
Vision-Language-Action (VLA) Model
Imitation Learning
Cross-Embodiment Learning
Data Pyramid
Real-World Robot Experiments

Chapters:
00:00 - Introduction
00:49 - GR00T N1 Overview
02:16 - The Full Stack Solution
03:08 - Focus on Humanoid Robots
04:39 - Foundation Models and Robotics
05:30 - Data Island Problem
06:12 - Cross Embodiment Support
06:32 - Dual System Architecture
07:35 - System 2 says, System 1 does
08:01 - The Data Pyramid
09:17 - Training Strategy
10:31 - Releasing the Model
10:55 - Model Architecture
11:33 - Eagle 2 VLM
12:26 - System 1
13:57 - Synthetic Data
14:37 - AI Generated Robot Manipulation
15:22 - DexMimogen
16:17 - Testing the Model
17:50 - Real Robot Tests
19:08 - Data Efficiency
20:09 - Limitations
20:48 - Related Work
22:37 - Conclusion