Bringing Robotics AI to Embedded Platforms: Dataset Recording, VLA Fine‑Tuning, and On‑Device Optimizations

This article presents NXP's best practices for deploying Vision-Language-Action (VLA) models on the i.MX 95 embedded SoC, covering dataset recording, fine-tuning, and on-device optimization.

• •Dataset quality prioritizes consistency via fixed cameras, controlled lighting, strong contrast, and calibration backups over sheer volume
• •Three cameras (top, gripper, left at 640×480px/30fps) are used; a gripper-mounted camera most effectively improves fine manipulation accuracy and enforces correct data collection
• •120-episode training sets span 11 workspace clusters with ~20% recovery episodes and a held-out validation cluster to prevent overfitting
• •SmolVLA graphs are decomposed into vision encoder, LLM backbone, and action expert blocks, enabling per-block quantization from 8-bit mixed precision to 4-bit depending on sensitivity