smolvla_libero_object_ft

SmolVLA Fine-tuned on LIBERO-Object

This model is a fine-tuned version of lerobot/smolvla_libero specifically trained on the LIBERO-Object task suite. It is designed to perform robotic manipulation tasks by taking visual observations and language instructions as input to predict 7-DoF actions.

🤖 Model Details

• Model Type: Vision-Language-Action (VLA)
• Base Model: lerobot/smolvla_libero (which is based on SmolVLM2)
• Task Suite: LIBERO-Object
• Framework: LeRobot
• Action Space: 7-DoF (dx, dy, dz, droll, dpitch, dyaw, gripper)
• Language: English (Task Instructions)

📊 Training Information

The model was trained by freezing the vision and language encoders and fine-tuning only the action expert modules and state projection layers.

• Dataset: lerobot/libero (filtered for LIBERO-Object tasks)
• Training Steps: 10,000 steps
• Batch Size: 8 (per device)
• Learning Rate: 1e-4 (Cosine decay with warmup)
• Hardware: Single NVIDIA T4 GPU
• Training Time: ~4 hours 50 minutes

🚀 How to Use

You can load and use this model directly with the Hugging Face lerobot library.

Installation

pip install lerobot