LiteVLA-Edge: Quantized On-Device Multimodal Control for Embedded Robotics

Researchers developed LiteVLA-Edge, a deployment-oriented Vision-Language-Action model pipeline that enables fully on-device inference on embedded robotics hardware like Jetson Orin. The system achieves 150.5ms latency (6.6Hz) through FP32 fine-tuning combined with 4-bit quantization and GPU-accelerated inference, operating entirely offline within a ROS 2 framework.

• →LiteVLA-Edge enables on-device multimodal AI control for embedded robotics without requiring cloud connectivity
• →The system achieves practical real-time performance with 150.5ms end-to-end latency on Jetson Orin hardware
• →Implementation uses 4-bit GGUF quantization and llama.cpp runtime for efficient GPU-accelerated inference
• →The solution integrates with ROS 2 framework maintaining modular interfaces between perception, reasoning, and actuation
• →Results establish timing feasibility for reactive language-conditioned robotic control in embedded environments