Driving · 2026-11-04
Aero-VPA: Few-Step Flow-Matching Control over Non-Reconstructive JEPA Latents for Agile Flight on Edge Hardware
Yeonseok Lee
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
Aero-VPA is a lightweight Vision-Predictor-Action architecture for agile 3D navigation on edge hardware, using a non-reconstructive JEPA latent world model plus a Conditional Flow Matching controller. The sampler generates a velocity/yaw-rate chunk in two Euler steps, requiring five sequential network evaluations per decision, with an ensemble-uncertainty fallback and a map-free runtime shield. The authors report first closed-loop PX4 SITL results in procedurally generated forests.
Why it matters
The paper targets end-to-end agile flight under embedded compute and fixed-rate control constraints, while addressing closed-loop failure modes of imitation learning and the need for safe actions when perception is unreliable. The reported gains come from a runtime shield that uses depth memory and control-barrier-style constraints to recover from blind decisions and reduce crashes/timeouts in closed-loop evaluation.
Method
- Architecture: compact RGB-D Vision Transformer encoder plus a five-head ensemble JEPA predictor trained without pixel reconstruction to forecast the next latent state; the ensemble mean conditions the actor and dispersion drives an uncertainty fallback.
- Control: Conditional Flow Matching sampler with K=2 Euler steps; it generates a velocity and yaw-rate chunk in two Euler steps so each decision uses five sequential network evaluations; added noise-persistence and inter-chunk handling via continuous blending.
- Safety: map-free runtime shield with a 15 s depth voxel memory, projection onto barrier constraints, a cap on motion outside the 73° field of view, and recovery/backoff behavior; trained/used with DAgger against a privileged, safety-gated supervisor.
Limitation
The paper states that the embedded deployment is specified but not yet built, and that timing was measured on an NVIDIA L4 GPU rather than the Jetson Orin Nano; it also notes that the hardware-in-the-loop stage and most ablations of the evaluation protocol have not been executed.
Abstract (from arXiv)
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.
Related papers
- DiffWAM: A Fast and Efficient Navigation World Action Model
- WareFly-VLA: A Vision-Language-Action Framework for UAV Navigation and Human Tracking in Smart Warehouses
- ScanDreamer: A Map-Grounded World Model for Quadrotor Navigation
- VeriVLA: Adaptive Test-Time Verification for Flow-Based VLAs with Latent World Models
- Retrieve to Act: Motion Primitive Graph Retrieval for Robust VLA Execution
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-11-03 · Indexed by Pub-AI: 2026-11-04 UTC
Aero-VPA: Few-Step Flow-Matching Control over Non-Reconstructive JEPA Latents for Agile Flight on Edge Hardware
Preprint posted: 2026-11-03 · Original record ↗
- arxiv
- 2611.04392
Explicit source versions (1)
- v1 · 2026-11-03
- Preprint available
Source date: 2026-11-03 · First detected: 2026-11-04 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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.
