X-G1

X-G1 is a rapid-iteration humanoid robotics pipeline built during a 2-day hackathon using the Unitree G1.
Our goal was to create a workflow that moves quickly from human teleoperation → dataset creation → model fine-tuning → autonomous policy evaluation.

By combining immersive teleoperation with scalable data infrastructure and embodied AI models, we demonstrate a fast-track workflow for humanoid robot learning.

---

Project Overview

In this project we built a complete pipeline for training and evaluating policies on the Unitree G1 humanoid robot.

The workflow includes:

1. Teleoperation\
2. Data storage & streaming\
3. Policy fine-tuning\
4. Diagnostics & failure analysis

This allows rapid experimentation with manipulation and locomotion tasks such as:

• Walking to tables
• Beverage organization
• Apple pick-and-place

---

System Architecture

Meta Quest 3 Teleoperation
        │
        ▼
MuJoCo Simulation + NVIDIA Sonic
        │
        ▼
Demonstration Data Collection
        │
        ▼
DeepLake Dataset Storage
        │
        ▼
GR00T Policy Fine-tuning
        │
        ▼
Autonomous Policy Evaluation
        │
        ▼
Nomadic AI Diagnostics

---

Technical Implementation

Teleoperation

We integrated Meta Quest 3 with MuJoCo using NVIDIA Sonic to enable low-latency humanoid control.

This setup allowed us to manually complete complex tasks including:

• Navigating the robot toward tables
• Picking and placing beverages
• Manipulating apples

The teleoperation pipeline enables fast demonstration collection for downstream training.

Teleop implementation:
https://github.com/AIBotTeachesAI/quest3-g1-teleop

---

Data Strategy (DeepLake)

To support fast training iteration, we used DeepLake for dataset storage and streaming.

We leveraged Lightwheel's G1 beverage organization dataset, storing the data in an optimized tensor format to enable:

• Fast random access
• Efficient streaming
• High-throughput training

This was critical for enabling model fine-tuning within the time constraints of the hackathon.

Data processing pipeline:
https://github.com/sl628/g1_hackathon/tree/main/data_processing

---

Policy Fine-Tuning (NVIDIA GR00T)

We fine-tuned NVIDIA GR00T N1.6 using the collected teleoperation data.

Because Sonic's fine-tuning features are not yet released, we used:

• Sonic → high-fidelity teleoperation and data collection
• GR00T → autonomous policy inference

Fine-tuning reference implementation:
https://github.com/NVIDIA/Isaac-GR00T/tree/main/examples/GR00T-WholeBodyControl

---

Diagnostics with Nomadic

To better understand policy failures, we integrated Nomadic AI as a diagnostic layer.

This allowed us to:

• Analyze incorrect task executions
• Inspect model reasoning
• Identify failure modes in task instructions

These insights help guide future improvements in policy robustness.

---

Key Contributions

• Rapid teleop → training → evaluation pipeline for humanoid robots
• Integration of VR teleoperation with MuJoCo simulation
• Efficient humanoid dataset handling using DeepLake
• GR00T-based policy fine-tuning
• Nomadic AI diagnostics for model reasoning analysis

---

Future Work

• Deploy policies directly on Unitree G1 hardware
• Expand demonstrations for broader manipulation tasks
• Integrate VLM-based instruction interfaces
• Improve policy robustness through larger datasets

---

Team

Team Name: X-G1
Hackathon Project (2 days)

---

Acknowledgements

We thank the teams behind:

• Unitree G1
• NVIDIA GR00T
• NVIDIA Sonic
• Lightwheel Robotics
• DeepLake
• Nomadic AI

for providing the tools that made this project possible.