Saved in:
Bibliographic Details
Main Authors: Williams, Justin, Gupta, Kishor Datta, George, Roy, Sarkar, Mrinmoy
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2603.03380
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912941336428544
author Williams, Justin
Gupta, Kishor Datta
George, Roy
Sarkar, Mrinmoy
author_facet Williams, Justin
Gupta, Kishor Datta
George, Roy
Sarkar, Mrinmoy
contents Vision-Language-Action (VLA) models provide a unified framework for perception, language conditioning, and action generation, but many existing systems remain difficult to deploy in embedded robotic settings because of their computational requirements and inference latency. In this paper, we present LiteVLA-Edge, a deployment-oriented VLA pipeline for fully on-device inference on Jetson Orin-class hardware. Our approach combines supervised image-to-action fine-tuning in FP32 with post-training 4-bit GGUF quantization and GPU-accelerated inference through the \texttt{llama.cpp} runtime. Under our deployment configuration, LiteVLA-Edge achieves a mean end-to-end latency of 150.5\,ms (approximately 6.6\,Hz) while operating entirely offline within a ROS~2-integrated perception--reasoning--action pipeline. Rather than introducing a new policy objective, our contribution is a practical systems path for executing compact multimodal control models locally on embedded hardware while preserving modular interfaces between perception, reasoning, and actuation. These results establish timing feasibility for reactive language-conditioned control and provide a reproducible baseline for future task-level evaluation of on-device VLAs in robotics.
format Preprint
id arxiv_https___arxiv_org_abs_2603_03380
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle LiteVLA-Edge: Quantized On-Device Multimodal Control for Embedded Robotics
Williams, Justin
Gupta, Kishor Datta
George, Roy
Sarkar, Mrinmoy
Robotics
Artificial Intelligence
Vision-Language-Action (VLA) models provide a unified framework for perception, language conditioning, and action generation, but many existing systems remain difficult to deploy in embedded robotic settings because of their computational requirements and inference latency. In this paper, we present LiteVLA-Edge, a deployment-oriented VLA pipeline for fully on-device inference on Jetson Orin-class hardware. Our approach combines supervised image-to-action fine-tuning in FP32 with post-training 4-bit GGUF quantization and GPU-accelerated inference through the \texttt{llama.cpp} runtime. Under our deployment configuration, LiteVLA-Edge achieves a mean end-to-end latency of 150.5\,ms (approximately 6.6\,Hz) while operating entirely offline within a ROS~2-integrated perception--reasoning--action pipeline. Rather than introducing a new policy objective, our contribution is a practical systems path for executing compact multimodal control models locally on embedded hardware while preserving modular interfaces between perception, reasoning, and actuation. These results establish timing feasibility for reactive language-conditioned control and provide a reproducible baseline for future task-level evaluation of on-device VLAs in robotics.
title LiteVLA-Edge: Quantized On-Device Multimodal Control for Embedded Robotics
topic Robotics
Artificial Intelligence
url https://arxiv.org/abs/2603.03380