theconstruct-ai/merged_dataset

Model Card for GR00T G1 - Merged Package Sorting VLA

Model Overview

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

Training Setup

Base Checkpoint

• Model: theconstruct-ai/sort_packages_real_4000
• Foundation Model: Pre-trained from nvidia/GR00T-N1.7-3B
• Original Data: Fine-tuned on the initial theconstruct-ai/sort_packages_real dataset.

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 & Catastrophic Forgetting Mitigation

To prevent catastrophic forgetting, the original dataset (theconstruct-ai/sort_packages_real) was deliberately re-included within the fine-tuning mixture (theconstruct-ai/merged_dataset). Training on a combined corpus of historical demonstrations and new episodes ensures the policy retains previously learned manipulation capabilities while adapting to new execution contexts and trajectory variations.

Training Hyperparameters

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

Technical Specifications

• Embodiment Tag: g1_29dof
• Action Representation: 29 DOF (Relative Joint & EEF control)