LUNA: Real-Time VLAs via Future-state-aware Asynchronous Inference

LUNA is an efficient and easy-to-use framework for VLAs inference.

Features

LUNA is efficient through:

• Asynchronous Inference: Overlaps action prediction with execution to achieve real-time performance
• Future-state Awareness: Uses predicted future states for more accurate action planning

LUNA is easy to use with:

• Simple CLI: Clean command-line interface for inference and serving
• Flexible Configuration: YAML-based configuration for easy customization

Installation

# Create conda environment
conda create -n "luna" python=3.10
conda activate luna

# Install dependencies
pip install -e .

Usage

# Run inference with a trained policy
luna run examples/inference/async.yaml

# Start model server for remote inference
luna server examples/inference/async.yaml

# Run inference with custom parameters
luna run examples/inference/async.yaml --action_quant_ratio=2

LUNA is designed to be flexible and extensible, allowing users to easily integrate new policies and customize existing ones.