Easy-to-use VLA deployment, fast to react, smooth in motion.

Paper

---

About

VLASH is an efficient and easy-to-use framework for VLAs fine-tuning and inference.

VLASH is efficient through:

• Asynchronous inference for fast reaction and smooth motion in real-time (>30Hz inference frequency for $\pi_{0.5}$ on RTX 5090)
• Future-state-awareness to enable stable asynchronous VLA inference without overhead
• Action quantization for faster robot execution speed
• LoRA with shared observation encoding for efficient fine-tuning on consumer GPUs

VLASH is easy to use with:

• Seamless integration with LeRobot datasets (v2.1, v3.0), models and robots
• Simple YAML-based configuration system
• Support for various policy architectures (e.g., $\pi_{0.5}$, $\pi_0$, ...)
• Easy deployment on real robot hardware

Demo

Getting Started

conda create -n "vlash" python=3.10
conda activate vlash
conda install ffmpeg=7.1.1 -c conda-forge
pip install -e .
pip install -U torch torchvision torchcodec

Quick Examples

Fine-tune a VLA policy for your task, enabling smooth async inference without overhead:

vlash train examples/train/pi05/async.yaml

Run async inference on a robot:

vlash run examples/inference/async.yaml

Run async inference with 2x speedup:

vlash run examples/inference/async.yaml --action_quant_ratio=2

TODO

• LoRA fine-tuning for $\pi_{0.5}$, $\pi_0$ under 12G GPU memory
• QLoRA fine-tuning for $\pi_{0.5}$, $\pi_0$ under 8G GPU memory
• Efficient fine-tuning with shared observation

Acknowledgment

This project is built upon the following excellent open-source projects: LeRobot, PEFT.

License

Apache 2.0