Aero-VPA: Few-Step Flow-Matching Control over Non-Reconstructive JEPA Latents for Agile Flight on Edge Hardware

Vision-Language-Action (VLA) models generalize across tasks, but their autoregressive decoders require a number of sequential forward passes that grows with the action horizon, and on size-, weight-, and power-constrained aerial platforms each pass is bounded below by the time needed to stream billions of parameters from memory. We present Aero-VPA, a lightweight Vision-Predictor-Action architecture for agile three-dimensional navigation on edge hardware. A compact Vision Transformer embeds RGB-D observations into a latent space shaped by a non-reconstructive Joint-Embedding Predictive Architecture (JEPA) world model, trained with variance--covariance regularization instead of pixel reconstruction. A discrete primitive selector and an ensemble latent predictor, conditioned on the active skill and the executing command, forecast the next latent state, and a Conditional Flow Matching (CFM) solver generates a complete chunk of velocity and yaw-rate setpoints in two Euler steps, so that every decision requires five sequential network evaluations irrespective of the horizon length. The normalized dispersion of the predictor ensemble provides an epistemic uncertainty signal that requires no future observation; it attenuates translational velocity, shortens the validity of each chunk, and triggers a minimum-jerk transition to hover, while a model-free braking envelope limits speed to the measured free space. We specify an asynchronous deployment on an NVIDIA Jetson Orin Nano (8\,GB, 15\,W) that streams interpolated setpoints to a PX4 flight controller over micro-XRCE-DDS, together with a software- and hardware-in-the-loop benchmark protocol against classical, reinforcement-learning, and generative baselines, with hypotheses, metrics, and ablations fixed in advance. No experimental results are reported.