kavinrajkrupsurge/openvla-phase-combined-finetuned

OpenVLA Fine-tuned Models - Phase Repository

Overview

This repository contains references to both fine-tuned OpenVLA models:

1. LAMP Search Model: kavinrajkrupsurge/lampe-sim-data-openvla
2. Keyboard Controlled Model: kavinrajkrupsurge/openvla-keyboard-controlled-finetuned

Models

LAMP Search Model

• Repository: kavinrajkrupsurge/lampe-sim-data-openvla
• Dataset: lampe_search_dataset/all
• Action Space: 4-DoF (Base, Joint2, Joint3, Joint4)
• Use Case: Search tasks in room environments

Keyboard Controlled Model

• Repository: kavinrajkrupsurge/openvla-keyboard-controlled-finetuned
• Datasets: keyboard_controlled_dataset/samples_000_024 + samples_025_049
• Action Space: 4-DoF (Base, Joint2, Joint3, Joint4)
• Use Case: General keyboard-controlled robot manipulation

Base Model

Both models are fine-tuned from: openvla/openvla-7b

Fine-tuning Configuration

• Method: LoRA (Low-Rank Adaptation)
• Batch Size: 4 (with gradient accumulation: 4)
• Effective Batch Size: 16
• Learning Rate: 5e-4
• LoRA Rank: 32
• LoRA Dropout: 0.0

Usage

Please refer to the individual model repositories for usage examples and loading instructions.

Training Scripts

Both model repositories include complete training scripts:

• LAMP Search: See LAMP Search Model Repo for training scripts
• Keyboard Controlled: See Keyboard Model Repo for training scripts

Each repository contains:

• Dataset conversion scripts
• Fine-tuning scripts
• Dataset configuration files
• Complete instructions for reproducing the training

Notes

• Both models use absolute joint positions (not deltas)
• Action normalization is handled automatically using dataset_statistics.json
• Each model has its own dataset statistics file
• All training was done with LoRA (Low-Rank Adaptation) for efficient fine-tuning