front_camera_only_model

GR00T-N1.5-3B LoRA Fine-tuned Model

This is a LoRA fine-tuned checkpoint of nvidia/GR00T-N1.5-3B trained on single front camera data.

Model Details

• Base Model: nvidia/GR00T-N1.5-3B
• Training Method: LoRA (Low-Rank Adaptation)
• Training Steps: 100,000
• Final Training Loss: 0.053

Training Configuration

LoRA Parameters

• Rank (r): 8
• Alpha: 16
• Dropout: 0.1
• Target Modules: to_q, to_k, to_v (attention layers only)
• Trainable Parameters: 1,638,400 (0.06% of total)

Training Parameters

• Batch Size: 2 per GPU
• Learning Rate: 1e-4
• Weight Decay: 1e-5
• Warmup Ratio: 0.05
• Optimizer: AdamW
• LR Scheduler: Cosine
• Training Duration: ~1h 52m (6719 seconds)
• Training Speed: 14.88 steps/second

Model Components Tuned

• LLM Backbone: ❌ Frozen
• Vision Tower: ❌ Frozen
• Action Head Projector: ✅ Tuned
• Diffusion Model: ✅ Tuned

Dataset

• Embodiment: SO-100 robot with single front camera
• Camera Resolution: 320x240
• FPS: 30
• Action Dimensions: 6 (5 DoF arm + 1 gripper)
• Action Horizon: 16 timesteps
• Video Backend: torchvision_av

Usage

This is a LoRA adapter that must be loaded on top of the base model:

from gr00t.model.gr00t_n1 import GR00T_N1_5
from peft import PeftModel

# Load base model
base_model = GR00T_N1_5.from_pretrained("nvidia/GR00T-N1.5-3B")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "path/to/this/checkpoint")

# Use for inference
model.eval()

Model Architecture

• Action Dimension: 32 (max)
• Action Horizon: 16
• Hidden Size: 2048
• Compute Dtype: bfloat16
• Diffusion Timesteps: 4 (inference)

Training Hardware

• GPUs: 1x NVIDIA GPU
• Compute Dtype: bfloat16
• TF32: Enabled
• Gradient Checkpointing: Disabled

Citation

If you use this model, please cite the original GR00T paper and model:

@misc{gr00t2024,
  title={GR00T: Generalist Robot Policy},
  author={NVIDIA},
  year={2024},
  url={https://huggingface.co/nvidia/GR00T-N1.5-3B}
}

License

Inherits license from nvidia/GR00T-N1.5-3B base model.