LiteVLA-Edge: 嵌入式机器人量化设备端多模态控制

Vision-Language-Action (VLA) models offer a unified framework for perception, language conditioning, and action generation. However, many existing systems face challenges in deployment within embedded robotic environments due to their computational demands and inference latency. LiteVLA-Edge presents a deployment-oriented VLA pipeline designed for fully on-device inference on Jetson Orin-class hardware. The approach integrates supervised image-to-action fine-tuning in FP32 with post-training 4-bit GGUF quantization and GPU-accelerated inference via the llama.cpp runtime. Under the specified deployment configuration, LiteVLA-Edge achieves a mean end-to-end latency of 150.5 ms (approximately 6.6 Hz) while operating entirely offline.