fine-tuning-openvla-oft

Fine-tune OpenVLA-OFT policies by customizing them for specific action generation needs using pre-defined workflows. This Skill streamlines the process with LoRA adaptation, continuous action heads, and FiLM conditioning for precise robot control.

Evaluate OpenVLA-OFT policies by assessing their performance directly on LIBERO and ALOHA simulation setups. The Skill provides built-in evaluation workflows to measure action generation accuracy and control efficiency.

Specific versions of PyTorch 2.2.0, Transformers 4.40.0, and TensorFlow 2.15.0 are required to fine-tune OpenVLA-OFT. These exact dependencies ensure computational reproducibility during policy training and evaluation.

OpenVLA-OFT fine-tuning supports LoRA adaptation and FiLM conditioning alongside continuous action heads. These features enable efficient policy customization and precise environmental control during robot action generation.

Fine-tune OpenVLA-OFT+ on ALOHA real-world data by cloning the repository and following the setup instructions. The Skill provides pre-defined workflows and computational requirements to streamline the training process.

OpenVLA-OFT fine-tuning requires PyTorch 2.2.0 and TensorFlow 2.15.0 to ensure full computational reproducibility. Maintaining both frameworks at specific versions guarantees consistent policy training and evaluation results across environments.