One lightweight AI brain for robot arms and drones

SLING AI builds VPA (Vision-Predictor-Action), control software that lets a machine look through its camera, predict the next moment and move, with no large language model in the loop. The same design steers a robot arm and a drone, and at about 70 MB it runs on the machine's own low-power board instead of a costly GPU.

Where we are: Working software in simulation, released openly as models, code and data. Tests on a real robot arm and a real drone begin in the first half of 2027.

Robot arm VPA

One VPA brain

  1. Visionsees
  2. Predictorpredicts
  3. Actionmoves

About 70 MB. Runs inside the device, no costly GPU.

Drone Aero-VPA

SLING AI at a glance

What we sell, to whom, and how far along we are.

What we build
AI control software for physical machines. VPA turns camera images into the machine's next 16 movements: joint commands for a robot arm (VPA) or speed and heading commands for a drone (Aero-VPA).
Who it is for
Factories, logistics centers and drone operators that need machines to react quickly; robot, drone and chip makers that want a small AI brain to ship inside their products.
How we make money
On-site software: a paid proof of concept, then an annual subscription priced at 20% of the labor cost it saves. Technology licensing: a royalty for every product shipped with VPA inside.
Stage
Working MVP in simulation, with models, code and benchmark data public. Pre-revenue; first revenue planned for 2027.
Validation
KRW 200 million credit guarantee from Korea Credit Guarantee Fund (KODIT), October 2026. One registered Korean patent.
Company
SLING AI Inc., founded February 2026 in Incheon, Republic of Korea. Founder and CEO: Dr. Yeonseok Lee.

Robot AI today is heavy, slow and built one machine at a time

Most new robot models first turn what the camera sees into language, then reason with a large language model (LLM) before they move.

  • A huge brain

    LLM-based robot AI carries 4 GB or more of weights and needs an expensive GPU.

  • Slow reactions

    It computes its answer one word at a time, so it responds late when an obstacle suddenly appears.

  • A separate AI for every machine

    Robot arms and drones each get their own AI, and a drone can rarely carry a large model at all.

  1. Camera Sees
  2. LLM, 4 GB+ Needs an expensive GPU Computes one word at a time This is where it slows down
  3. Robot or drone Moves Reacts late
In the usual pipeline, turning what the machine sees into words is the heavy, slow step.

Slow, heavy and fragmented: this is what keeps robots and drones out of everyday field work.

How VPA works: see, predict, move

VPA drops the language step. Four small networks run in a row, and the last one writes the whole next stretch of motion in a single pass.

Camera Same input for robot arms and drones
1. See (Vision) Keeps only what matters in the image
2. Choose (Skill selector) Picks what to do next
3. Preview (Predictor) Predicts the scene 0.1 s ahead
4. Move (Action) Writes 16 actions at once
Next 16 actions in 0.012 s Robot-arm joints or drone flight

No language step: 5 network calls plan the whole continuous motion.

Key numbers

Time to plan the next 16 actions
12 ms, about a tenth of a blinkVPA measured on an Apple M4 chip
Model size
About 70 MB, roughly the size of 15 songs18.13 million parameters (Aero-VPA)
Compared with a 4 GB LLM-based robot AI
About 1/50 the size
Network calls per decision
Always 5, however long the motion
Where it runs
On the machine's own low-power boardAero-VPA is designed for the 15 W NVIDIA Jetson Orin Nano
Where we train it
Google Cloud GPU virtual machines (NVIDIA L4)The drone flight simulation (PX4 and Gazebo) runs there too

Example: a drone flying through a forest to its goal

1. See (V)
The camera locates the trees ahead.
2. Choose
It picks “avoid the trees and keep moving forward”.
3. Preview (P)
It predicts the next 0.1 s to spot where it could hit something.
4. Move (A)
It sends 16 speed and heading commands at once.
Safety shield
It remembers passed trees for 15 s, and slows down or backs off when it gets too close.

70 MB refers to Aero-VPA's 18.13 million parameters stored at full precision. VPA for robot arms is the same size class, under 20 million parameters.

One VPA, two bodies

The same architecture becomes VPA when trained on robot-arm data and Aero-VPA when trained on drone data.

VPA for robot arms

Tested in simulation
What the robot sees: room camera (left) and wrist camera (right). Task: put both moka pots on the stove.

Success rate on 10 household tasks20 trials per task, first results

1 camera 10.5%
2 cameras 40.0%, about 4×

Aero-VPA for drones

Tested in simulated flights
Top view of a forest: the drone's camera view, the path flown so far with breadcrumbs, two remembered trees and the remaining route to the goal 1 2 3 Camera view 15 s memory Start Goal
Seen from above: the camera's view, the path flown so far and trees kept in memory for 15 s. The map is only for you; the drone flies from its camera alone.

Flights that reach the goal in a simulated forest180 flights; crashes fell from 72% to 22%

No safety shield 21.1%
With safety shield 52.2%, 2.5×

One VPA: the same architecture, about 70 MB, 5 network calls per decision, low-power boards.

Results so far come from simulation. Tests on a real robot arm and a real drone are scheduled for the first half of 2027.

How VPA compares

Robot AI today either runs a large language model or, for drones, builds a map before it can plan. VPA does neither.

Approach Model size (parameters) Hardware needed How it produces motion Used for
Large robot AIGoogle and other big-tech labs; LLM-based 3.3 billion to 55 billion High-end GPU Turns what it sees into words, then reasons Robot arms
Small robot AIOpen small models; LLM-based About 450 million Standard GPU or PC Turns what it sees into words, then reasons Robot arms
Map-based autonomous flightTraditional approach: builds a map first Not applicable Several sensors Builds a map, then computes a route Drones
SLING AIVPA and Aero-VPA About 18 million Low-power edge board 5 network calls, straight to motion Robot arms and drones, same architecture
  • Ultra-light

    1/100 the size of large robot AI, or smaller.

  • Fast reactions

    Every decision takes exactly 5 network calls.

  • One brain, two bodies

    The same architecture controls robot arms and drones.

  • Safety built in

    The safety shield cut drone crashes from 72% to 22%.

Model size is the number of parameters reported in public papers: Google RT-2 up to 55 billion, OpenVLA 7 billion, π0 3.3 billion, SmolVLA 450 million.

Business model

Two revenue streams, both from the same VPA software.

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

Why now

Demand for robots and drones is rising fast, and the cost of high-end GPUs is what holds deployment back.

Market

  • WorldwideRobotics and AI: KRW 250 trillion. Drones: USD 59.2 billion (2032 forecast).
  • Korea, where we startSmart factories and robots: KRW 11 trillion. Drones: KRW 1.1 trillion (2023).
  • Our first targetsRobots for logistics and precision manufacturing: KRW 1 trillion. Drone software: KRW 163.4 billion (2023).

One VPA lets us go after both markets at the same time.

Sources: Korea Ministry of Land, Infrastructure and Transport, 2024 Drone Industry Survey; IMARC.

Traction so far

Public validation
KRW 200 million credit guarantee from Korea Credit Guarantee Fund (KODIT)Selected for the Innovative Startup Partnership Guarantee, October 2026
Simulation results
Robot arm: 40% success on household tasks. Drone: 52% of flights reach the goal and 78% fly without a crash.
Open models
2 AI models, with their code and data, free to download Hugging Face GitHub
Papers
3 preprints on Zenodo; the VPA paper is also on arXiv Aero-VPA VPA NS-JEPA arXiv
Patents
1 registered Korean patent, plus several applications filed, including VPA's core technology Patent record

Recent updates

  1. Oct 2026Selected by KODIT for the Innovative Startup Partnership Guarantee (KRW 200 million).
  2. Oct 2026Aero-VPA preprint published on Zenodo. Its model weights and full benchmark data are public on Hugging Face.
  3. Oct 2026VPA preprint published on Zenodo and arXiv (2610.15556), with LIBERO-10 checkpoints on Hugging Face.
  4. Sep 2026NS-JEPA preprint published on Zenodo.
  5. Feb 2026SLING AI Inc. founded in Incheon, Republic of Korea.

Milestones

  1. 2026Done

    • Simulation validation and full open release
    • KODIT credit guarantee, KRW 200 million
  2. First half of 2027

    • Tests on a real robot arm and drone
    • 1–2 paid proofs of concept
  3. Second half of 2027

    • Field deployment and first license contract
    • Government R&D program (TIPS)
  4. 2028

    • Expansion to robot and drone makers
    • Series A investment

Team

AI control expertise grounded in mathematical proof.

Portrait of Dr. Yeonseok Lee

Dr. Yeonseok Lee

Founder and CEO

Ph.D. in Informatics, Nagoya University (2026)

  • Specialty: proving with mathematics that a program runs without errors (formal verification)
  • 3 peer-reviewed papers at international venues
  • Designed and built VPA and Aero-VPA, as sole author of the papers and the code
  • Hanyang University; Republic of Korea Air Force aircraft weapons maintenance; NTT DATA group, Japan

Full profile and publications

“Innovating AI control by combining theoretical rigor with hands-on field experience.”

Company and contact

Legal name
SLING AI Inc. (슬링에이아이 주식회사)
CEO
Dr. Yeonseok Lee
Founded
February 9, 2026
Business registration no.
399-88-03360
Corporate registration no.
120111-0151637
Address
Room 672, 5-6F, 185 Incheon Tower-daero, Yeonsu-gu, Incheon 22009, Republic of KoreaFASTFIVE Songdo Central
Email
ylee@sling.ai.kr
Phone
+82 10-2158-9243

Full company information