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

DevFeed: [Aero-VPA: Few-Step Flow-Matching Control over Non-Reconstructive JEPA Latents for Agile Flight on Edge Hardware](<https://devfeed.tech/articles/aero-vpa-few-step-flow-matching-control-over-non-reconstructive-jepa-latents-for-agile-flight-on-edge-hardware-116937.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.04392>)

Author: Yeonseok Lee

Published: 2026-11-04T05:00:00Z

Content type: news

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Physical AI](<https://devfeed.tech/topics/physical-ai.md>), [Embedded Systems](<https://devfeed.tech/topics/embedded-systems.md>), [Embedded Software Dev](<https://devfeed.tech/topics/embedded-software-dev.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [3d](<https://devfeed.tech/tags/3d.md>), [agile](<https://devfeed.tech/tags/agile.md>), [autoregressive](<https://devfeed.tech/tags/autoregressive.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [benchmark-dataset](<https://devfeed.tech/tags/benchmark-dataset.md>), [code](<https://devfeed.tech/tags/code.md>), [command](<https://devfeed.tech/tags/command.md>), [conditional](<https://devfeed.tech/tags/conditional.md>), [control](<https://devfeed.tech/tags/control.md>), [crashes](<https://devfeed.tech/tags/crashes.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-lo](<https://devfeed.tech/tags/cs-lo.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [dagger](<https://devfeed.tech/tags/dagger.md>), [deployment](<https://devfeed.tech/tags/deployment.md>), [dispersion](<https://devfeed.tech/tags/dispersion.md>), [edge](<https://devfeed.tech/tags/edge.md>), [embedded](<https://devfeed.tech/tags/embedded.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [flight](<https://devfeed.tech/tags/flight.md>), [flow-matching](<https://devfeed.tech/tags/flow-matching.md>), [hardware](<https://devfeed.tech/tags/hardware.md>)

## AI overview

Aero-VPA combines a compact RGB-D Vision Transformer, a five-head predictor trained without pixel reconstruction, and a Conditional Flow Matching controller that produces velocity and yaw-rate chunks in two Euler steps, using five sequential network evaluations per decision. In PX4 software-in-the-loop forest missions, behaviour cloning failed on all 60 missions, while DAgger policies reached 11.1-21.1% success. Adding a map-free runtime shield with depth memory, barrier constraints, a field-of-view motion cap, and path backoff raised the first DAgger policy's success from 21.1% to 52.2% over 180 paired missions and reduced crashes from 72.2% to 22.2%; timeouts increased from 2.8% to 23.3%. All six checkpoints gained 30.0-34.4 percentage points. These are simulation results; the excerpt does not report real-flight evaluation.

## Source excerpt

arXiv:2611.04392v1 Announce Type: cross Abstract: Vision-Language-Action (VLA) models generalize across tasks, but their autoregressive decoders need sequential forward passes in proportion to the action horizon, each bounded below on embedded aerial platforms by streaming billions of parameters from memory. We present Aero-VPA, a lightweight Vision-Predictor-Action architecture for 3D navigation on edge hardware. A compact RGB-D Vision Transformer and a five-head ensemble predictor, trained without pixel reconstruction, forecast the next latent state; a Conditional Flow Matching solver generates a velocity and yaw-rate chunk in two Euler steps, so each decision takes five sequential network evaluations; and the ensemble dispersion drives an uncertainty fallback. Version 1 specified the architecture, deployment and protocol without results. This version reports the first closed-loop evaluation: PX4 software-in-the-loop simulation of quadrotor missions in procedurally generated forests. Behaviour cloning failed on all 60 missions despite open-loop errors of 0.02-0.04 m/s: the policy copies its last command, a copycat form of causal confusion. DAgger with a privileged, safety-gated supervisor gave the first successes but plateaued at 11.1-21.1% across six checkpoints. A map-free runtime shield with a 15 s depth memory, barrier constraints, a cap on motion outside the 73-degree field of view and backoff along the flown path raised the first DAgger policy's success on 180 paired missions from 21.1% to 52.2% (95% CI [45.0, 59.4]; exact McNemar p = 1.2 x 10^-10) and cut its crashes from 72.2% to 22.2%, while timeouts rose from 2.8% to 23.3%. All six checkpoints gained 30.0-34.4 percentage points, at mean speeds of 1.16-1.36 m/s. The code (https://github.com/ylee-sling/Aero-VPA), the trained policy (72.6 MB, https://huggingface.co/ylee-sling/aero-vpa) and the full benchmark dataset are public.