Reproduce “LiteVLA-Edge: Quantized On-Device Multimodal Control for Embedded Robotics” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Justin Williams, Kishor Datta Gupta, Roy George, Mrinmoy Sarkar
LiteVLA-Edge presents a system for real-time, on-device multimodal robotic control, leveraging aggressive 4-bit quantization and optimized GPU inference on NVIDIA Jetson AGX Orin hardware. It achieves a mean end-to-end inference latency of 150.5 ms (6.6 Hz), enabling reactive visuomotor control for embedded robotics in resource-constrained environments.
Key insights
- Aggressive 4-bit quantization can be effectively applied to compact VLA models for robotic control while maintaining sufficient representational density for stable and precise motor command generation.
- Full GPU offloading of all transformer layers and optimized C++ kernels for quantized inference can dramatically reduce VLA inference latency on embedded GPUs, enabling a shift from deliberative to reactive control.
- A modular VLA design allows for real-time, on-device image-to-action generation, preserving clear interfaces for debugging and safety overrides, which is critical for robust robotic systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.