Summary VLA-Adapter offers an efficient paradigm for tiny-scale Vision-Language-Action (VLA) models in robotics. It enables adaptation of large Vision-Language Models (VLMs) for embodied tasks, benefiting researchers and developers with limited computational resources.

How It Works This project implements a lightweight adapter module bridging pre-trained VLMs with robotic control policies. It facilitates efficient fine-tuning on datasets like LIBERO and CALVIN by leveraging VLM backbones (e.g., Prismatic-VLMs Qwen2.5-0.5B) for embodied tasks, prioritizing efficiency and adaptability for resource-constrained environments.

Quick Start & Requirements

• Installation: Requires Python 3.10.16, PyTorch 2.2.0, and Flash Attention 2. Setup involves creating a Conda environment, cloning the repo, and running pip install -e . and flash-attn.
• Dependencies: Requires separate installation and downloading of LIBERO (approx. 10GB) and CALVIN (approx. 50GB) benchmark datasets and their repositories. A VLM backbone (e.g., Prismatic-VLMs Qwen2.5-0.5B) must be downloaded and placed in /pretrained_models.
• Hardware: Training supports GPUs from 10GB VRAM (requiring careful parameter tuning) up to multi-GPU setups (e.g., 4x H100). Inference can be performed on single GPUs.
• Links: Paper: https://arxiv.org/pdf/2509.09372, Project Page: https://vla-adapter.github.io/, HuggingFace Models: https://huggingface.co/VLA-Adapter.

Highlighted Details

• Achieves state-of-the-art performance on the LIBERO benchmark for tiny-scale VLA models (Pro version: 98.5% avg success).
• Demonstrates strong performance on the CALVIN benchmark (Pro version: 80.0% avg success).
• Provides detailed training guidance for GPUs with VRAM as low as 10GB, enhancing accessibility.
• Offers an original and an enhanced "Pro" version with improved performance and smaller model size.

Maintenance & Community The project recently released its code and paper, with ongoing development planned for enhanced versions (VLA-Adapter++), broader system compatibility, and integration with more foundation models. Community links include Twitter and WeChat.

Licensing & Compatibility The README does not explicitly state a license. Clarification is needed for commercial use or closed-source linking, though the project builds on other open-source works.

Limitations & Caveats The project is under active development with planned improvements. Performance may vary slightly on GPUs other than NVIDIA H100 for inference. No specific known bugs or deprecations are mentioned.