Status: Research prototype tested in simulated flights (PX4 SITL with Gazebo Harmonic). Code, trained weights and every benchmark record are public; flights on a real drone are planned for the first half of 2027.
How it works
1. World model
A Vision Transformer embeds each RGB-D camera frame (160 × 128). A JEPA predictor, trained to predict the next embedding rather than pixels, forecasts the scene 0.1 s ahead; the disagreement of its five heads measures uncertainty.
2. Actor
A primitive selector and a conditional flow-matching network turn the latents, the next waypoint and the vehicle state into a 16-step velocity and yaw-rate chunk (0.32 s) in two integration steps. A decision is five network calls, whatever the chunk length.
3. Safety shield
A deterministic runtime filter keeps 15 s of depth returns in a local 3-D memory, so trees beside and behind the 73° camera stay known. It limits the approach speed to every remembered obstacle, caps motion outside the field of view and backs off along the flown path when too close. It uses no map.
The world model and the actor are the same Vision-Predictor-Action design as VPA for robot arms: no language decoder in the control loop, so the cost of a decision does not grow with the length of the motion.
Results in simulation
- Goal reached, recommended policy (r1)
- 52.2% with the shield, 21.1% without94 of 180 missions, 95% CI 45.0–59.4%; paired exact McNemar p = 1.2 × 10⁻¹⁰.
- Goal reached, all six policies
- 42.8–52.2% with the shieldEvery checkpoint improved significantly with the shield.
- Crashes
- From 72–82% down to 22–43%Across the six checkpoints; r1 goes from 72% to 22%, i.e. 78% of its flights end without a crash.
- Test setup
- 36 test worlds × 5 missionsForest worlds of three densities, disjoint from training; every policy flies the same missions.
- Model size
- 18.13 million parametersAbout 36 MB at FP16 (72 MB at FP32).
- Control rate
- 100 Hz velocity setpoints to PX4Asynchronous two-rate stack designed for a 15 W NVIDIA Jetson Orin Nano.
PX4 SITL with Gazebo Harmonic. The model card on Hugging Face has the per-checkpoint and per-density tables, and sim/paired_stats.py in the code repository recomputes every number from the published records. These are simulation results with one training seed per recipe.
Parameter budget
| Module | Parameters | Role |
|---|---|---|
| Vision encoder | 11.17M | 4-channel RGB-D (128 × 160) Vision Transformer to a 256-dim latent |
| Primitive selector | 0.67M | Chooses one of 6 flight primitives |
| JEPA predictor | 2.72M | 5-head ensemble forecasting the next latent and its uncertainty |
| Flow-matching solver | 3.57M | 16-knot, 4-DoF heading-frame velocity chunk |
| Total | 18.13M | About 36.3 MB of FP16 weights |
How customers use it
On-site software
- Customers
- Factories, logistics centers and drone operators
- Price
- 20% of the labor cost the software saves
- Sales path
- Paid proof of concept (PoC), then an annual subscription
Technology licensing
- Customers
- Robot, drone and chip makers
- Price
- Royalty per unit shipped
- Sales path
- The free open release builds adoption; products sold with VPA inside need a paid license
Open release and licensing
- Paper: “Aero-VPA: Few-Step Flow-Matching Control over Non-Reconstructive JEPA Latents for Agile Flight on Edge Hardware”, October 2026: Zenodo (CC BY 4.0)
- Code: github.com/ylee-sling/Aero-VPA (GPL-3.0-only)
- Weights: huggingface.co/ylee-sling/aero-vpa (GPL-3.0), six DAgger checkpoints with FPV flight videos
- Benchmark data: huggingface.co/datasets/ylee-sling/aero-vpa-benchmark (CC BY 4.0): every mission record, decision traces and the 36 test worlds
For commercial integration, proprietary licensing or defense adaptation, contact SLING AI.