smolvla-omy-checkpoints

SmolVLA-OMY Model Checkpoints

This repository contains training checkpoints for a SmolVLA (Small Vision-Language-Action) model trained on the ArrangeVegetables task.

Model Details

• Model Type: SmolVLA (Vision-Language-Action model)
• Task: ArrangeVegetables manipulation task
• Training Steps: 20,000 steps
• Batch Size: 350
• Chunk Size: 5 action steps
• Input Features:
  • Visual observations: 256x256 RGB images (both main camera and wrist camera)
  • State observations: 6-dimensional state vector
• Output Features: 12-dimensional action space

Checkpoint Structure

The repository contains checkpoints saved at different training steps:

• 000500/: Checkpoint at 500 steps
• 001000/: Checkpoint at 1,000 steps
• 001500/: Checkpoint at 1,500 steps
• 002000/: Checkpoint at 2,000 steps

Each checkpoint contains:

• pretrained_model/: Model weights and configuration
• training_state/: Optimizer state, scheduler state, and training metadata

Training Configuration

• Device: CUDA
• Seed: 42
• Workers: 24
• Evaluation Frequency: Every 5 steps
• Logging Frequency: Every step
• Image Resize: 512x512 with padding
• Normalization: Identity for visual, mean-std for state/action

Usage

To load a checkpoint:

from your_training_framework import load_checkpoint

# Load the latest checkpoint (2000 steps)
model = load_checkpoint("./002000/pretrained_model/")

Dataset

Trained on the ArrangeVegetables dataset available at: lava8888/ArrangeVegetables