theconstruct-ai/raw-merged-1000

Model Card for GR00T G1 - Direct Package Sorting VLA

Model Overview

Fine-tuned Vision-Language-Action (VLA) model based directly on NVIDIA GR00T for the Unitree G1 (23 DOF) humanoid robot, optimized for autonomous bimanual package sorting tasks.

Training Setup

Base Checkpoint

• Model: nvidia/GR00T-N1.7-3B
• Foundation Model: Pre-trained nvidia/GR00T-N1.7-3B (Direct fine-tuning, no intermediate checkpoints)

Training Dataset

• Dataset: theconstruct-ai/merged_dataset
• Composition:
  • theconstruct-ai/sort_packages_real
  • theconstruct-ai/sort-pckg-real-guillem
  • Newly collected real-world teleoperated demonstration episodes

Training Strategy

The policy was fine-tuned directly from the base foundation model using a comprehensive combined dataset (theconstruct-ai/merged_dataset). By training on a diverse mixture of historical demonstrations and newly collected teleoperated episodes simultaneously, the model learns robust manipulation capabilities and adapts to trajectory variations natively, without relying on or being biased by prior task-specific checkpoints.

Training Hyperparameters

• Training Steps: 1,000
• Hardware: 2x H200 SXM

Technical Specifications

• Embodiment Tag: g1_23dof
• Action Representation: 23 DOF (Relative Joint & EEF control)