Fly a real-feeling FPV drone through photoreal 3D Gaussian Splatting scans, right in your browser. ShramkoGSFPV loads scenes straight from SuperSplat, flies them with a quadcopter model that has four motors, a mixer, a rate PID and Betaflight-compatible rate curves, crashes you into real walls with swept voxel collision, and takes your own radio (EdgeTX over WebHID), a gamepad or touch sticks. WebGPU, the MIT-licensed PlayCanvas engine, nothing to install. Created by Andrii Shramko.
▶ Fly now: https://gsfpv.flyreelstudio.eu
Status: alpha, live since September 2026. The site, the simulator and the measurements below run in public. Every feature claim in this README is either measured (a JSON file in
evidence/with its method) or explicitly marked not tested yet. Most pass/fail checks also run a negative control, a deliberately broken case that must fail; where a check has none, or a weak one, the line says so. The radio path has been tested with a simulated EdgeTX radio, not yet with a real one.
Real screenshots of the simulator, taken on 27 September 2026 by tools/bench/src/screens.ts in system Chrome (1920×1080 at device scale 2, so the menus stay sharp). A menu is shown as itself with a little of the scene around it; click any picture for the whole screen. The calibration screens use the simulator's built-in simulated EdgeTX radio; the flights are flown by its test pilot and frozen while it is really flying. All 30 screens, phone-sized ones included, are in docs/screenshots/ and on the landing page.
 Voxel grid over the scan — the walls you crash into, drawn over the scan in height colours; `V` cycles off / over the scan / voxels only. |
 Voxels only — the scan hidden, its 5 cm walls as solid cubes: fly the grid alone and see every surface you can hit. |
 Flight view — Andrii's Tunis scan at 4.8 m/s with the OSD: arm state, flight mode, voltage, flight time, throttle, speed, altitude. |
 Calibration — a drawn radio shows which stick to move; here the throttle is found on CH1. |
 Pause menu — scan, drone, controls, settings, replays, measurements, Betaflight import, cinema mode, with their keys. |
 Settings — FOV, uptilt, HUD, walls and voxel grid, quality, gravity, crash threshold, motor spin-up, drag, PID. |
 Six drones — every number labelled manufacturer, measured or estimate, with the rate curve. |
 Betaflight import — paste a CLI diff; unused settings are listed (here a test diff). |
 Crash — impact speed, Rapier debris, respawn, a 10-second replay or the flight log. |
 Walls on or off — the walls line opens the walls (collisions) switch and the voxel grid; here the walls are off (`C`). |
 Keyboard — every flying key on screen, `V` voxel grid and `C` walls included. |
 Cinema mode — full detail, no HUD, MP4 recording with the scan's credit. |
 Wireframe grid — the Winter Garden's 3.2 cm walls as a wireframe over the photo scan. |
 Floaters in red — pieces that touch nothing else and are under 0.5 m³ turn red, so a phantom wall stands out. |
 Winter Garden — the test pilot's dash at 8.1 m/s between the table and the plants. |
 Choose a scan — Andrii's scans, recent and favourites, filters, or any SuperSplat link. |
FPV pilots train in simulators like Liftoff or VelociDrone because the physics and stick response feel right, but those sims only fly the maps someone built for them. Meanwhile people publish thousands of real places as 3D Gaussian Splatting scans — rooms, villas, streets, whole city blocks — and you can only walk or orbit through them.
ShramkoGSFPV joins the two: pick a real scanned place, arm, and fly it with a drone that behaves like the one in your hands. Rehearse an indoor line before flying it for real, plan a cinematic shot in the actual location, or fly your own house on the Moon.
The default craft is a BetaFPV Pavo20 Pro class 2.2″ cinewhoop, with five more presets: Pavo20 Pro II on 3S and 4S, Pavo Pico, Meteor65 Pro and Air65. Most scans are rooms and streets, where a 5″ freestyle quad has no room to move. On each preset card the spec rows (thrust-to-weight, weight, wheelbase, motor, battery) say where the number comes from: manufacturer, independent measurement or estimate. The card's maximum rate (°/s) and hover-throttle figures do not carry a source label yet. The thrust-to-weight of three presets (Pavo20 Pro 3S, Pavo20 Pro II 3S, Pavo Pico 2S) was measured in the model and is within 0.4 % of the preset value (b-fly-b9.json); the other three are not measured yet.
Status on 27 September 2026. Live means you can use it on https://gsfpv.flyreelstudio.eu today; In progress is being built now; Next is designed and queued. How each input device was tested is in Radios and devices, and the numbers behind the physics are in What is measured. The same list, in four languages, is on the landing page.
| Area | Feature | Status |
|---|---|---|
| Radio & calibration | EdgeTX radio over USB (WebHID) in Chrome or Edge | Live |
| Radio & calibration | Calibration with a drawn radio that shows which stick to move | Live |
| Radio & calibration | Reverse any channel after calibration | Live |
| Radio & calibration | Recalibrate button in Controls | Live |
| Radio & calibration | A known radio reconnects by itself when you change scans | Live |
| Gamepad, touch & keyboard | Gamepad through the browser's Gamepad API (not tested on real hardware yet) | Live |
| Gamepad, touch & keyboard | Touch sticks and an ARM button on a touch screen (tested in touch emulation only) | Live |
| Gamepad, touch & keyboard | Keyboard, with every flying key shown on screen | Live |
| Gamepad, touch & keyboard | Keyboard: M switches ANGLE (self-levelling) and ACRO; touch flies in ANGLE, a radio or gamepad in ACRO (ANGLE not tested yet) |
Live |
| Gamepad, touch & keyboard | Close Controls with × or Esc; the pause menu shows its keys (Esc / P continue, R restart) | Live |
| Drone & tuning | Six BetaFPV-class presets: Pavo20 Pro 3S (default), Pavo20 Pro II 3S and 4S, Pavo Pico 2S, Meteor65 Pro 1S, Air65 1S | Live |
| Drone & tuning | Betaflight, Actual, KISS and Raceflight rate curves (set through the diff import), checked against Betaflight 4.5.1 (a3-rates-vectors.json) |
Live |
| Drone & tuning | Import rates, PID and throttle curve from a Betaflight CLI diff / diff all (d-bf-diff-import.json) |
Live |
| Drone & tuning | PID for each axis: P, I, D and F | Live |
| Drone & tuning | Motor spin-up time 8–30 ms and air drag 0.5–2× | Live |
| Gravity & crashes | Gravity: Earth and Moon (measured, b-fly-b16.json), Mars, zero-g and a custom value through the ?g= link parameter (not tested yet; a custom value matches no entry of the Settings gravity list, so Apply in Settings then sets zero-g) |
Live |
| Gravity & crashes | On other worlds: same motors, keep thrust-to-weight or auto throttle (the last two not tested yet) | Live |
| Gravity & crashes | Crash threshold from 2 to 10 m/s | Live |
| Gravity & crashes | After a crash: respawn at the start (b-fly-b12.json). The "From the last safe point" button does not keep its point reliably yet: the point is sampled only when a frame ends exactly on a 500th physics tick, so on a steady display it often stays at the start (defect D-d in docs/architecture-v03.md) |
Live |
| Gravity & crashes | Replay the last 10 s after a crash; a list of replays (b-fly-b15.json) |
Live |
| Camera, view & video | Camera field of view 70–150° and uptilt 0–50° | Live |
| Camera, view & video | Stable frame ↔ detail slider; quality steps down by itself when frames drop (d-governor.json) |
Live |
| Camera, view & video | HUD on or off; F3 shows frame times | Live |
| Camera, view & video | Reduced motion: a calmer crash camera (not tested yet) | Live |
| Camera, view & video | Cinema mode: full detail and MP4 recording on the showcase scans (d-cinema.json) |
Live |
| Camera, view & video | Simulator in English, Spanish, Polish or Russian | Live |
| Scenes & flight data | Any public SuperSplat scene in the current format, by link or id, re-published versions (v2, v3…) too; scenes in SuperSplat's older PLY format do not open yet | Live |
| Scenes & flight data | Tabs: Andrii's scans, Recent, Favourites; filters for walls, type and flown | Live |
| Scenes & flight data | Walls built in your browser for scans published without them, up to 4 M Gaussians (c-bake.json) |
Live |
| Scenes & flight data | Loading progress: stage, MB, speed, time left and retry | Live |
| Scenes & flight data | Save the flight log; export the trajectory as CSV or JSON (CSV checked in d-trajectory-export.json, JSON not tested yet) |
Live |
| Scenes & flight data | Measurements panel: display refresh, frame times, physics rate and the input in use (its report rate is not measured yet); copy the report as JSON | Live |
| Scenes & flight data | Voxel grid over the scan or on its own with the scan hidden (V): solid cubes, wireframe, height colours or floaters in red (pieces under 0.5 m³ that touch nothing), opacity per mode; test switches ?voxels=overlay|only&vstyle=&vopacity=. Faces and floaters checked against brute force with controls (packages/collision/test/); frame cost measured only on a 30 Hz screen, no evidence file yet |
Live |
| Scenes & flight data | Walls (collisions) on or off (C, the walls menu or Settings), remembered per scan in the browser; the admin default per scan is "walls": "on" | "off" in apps/fly/public/showcase.json (off for a noisy scan). A flight log records the setting and replays with it; on hits the wall, off flies through it, each log replays to its own hash, with controls (walls-switch.test.ts) |
Live |
Built on one shared settings store and one respawn system, not as quick patches (docs/architecture-v03.md).
| Feature | Status |
|---|---|
| Calibration that waits for you: a step moves on only when you really move the stick | In progress |
| Every setting kept between visits, with export, import and reset — one or all | In progress |
| Pick the flight mode with a click; self-levelling by default for everyone | In progress |
| Start on an invisible pad at spawn height instead of dropping to the floor | In progress |
| After a crash, back in the air within 2 s, on a pad 5 s earlier; R returns to the start | In progress |
| Crashes off for beginners | In progress |
| Automatic restart when you are stuck or lying upside down | In progress |
| Pavo20 Pro and Pro II physics tuned against the real quads | In progress |
| Feature | Status |
|---|---|
| Flight stats at the end of a flight, like the goggles OSD | Next |
| A full catalogue of every setting, on the site and here on GitHub | Next |
| After a crash: your next favourite or a random top-rated scene (auto-switch off by default) | Next |
| A SuperSplat tab with the same filters as superspl.at | Next |
| Scale a wrong-size scene around the drone | Next |
| 60 fps recording, and auto-record into the last folder you chose | Next |
Missing a setting you need? Suggest it through the contact form or open an issue.
Numbers from the JSON files in evidence/2026-09-24/ named in each row. The landing page's figures come from evidence/latest.json (rates, tunnelling, clearance, latency), and its build fails if a number is missing.
| Check | Result | How |
|---|---|---|
| Rate curves vs Betaflight 4.5.1 | 1880 points, max error 0 °/s in double precision | Betaflight's own rc.c compiled outside the repo with float promoted to double, vs our re-implementation; a float32 build differs by at most 6.3e-4 °/s (PROVENANCE). A deliberately wrong curve (x³ instead of x⁵ in Actual) misses the vectors by up to 280 °/s, so they tell the curves apart (a3-rates-vectors.json) |
| Flying through walls (tunnelling) | 0 in 200,000 straight passes | 1 ms swept test vs an independent resampling oracle at 5–34 m/s on 3 real scans and a 2 cm wall; the endpoint-only control misses walls, so the oracle is not blind (a5-tunnelling.json) |
| Determinism | same SHA-256 of the whole trace in Node and Chrome and at 30/60/144/240 Hz frame splits | Math.random injected into the core changes the hash (a6-determinism.json) |
| Replay from stick inputs only | same hash in a new tab; 1 LSB changed in one report → different hash and > 1 cm divergence | 30 s flight with a flip and a crash, on the live site (b-fly-b15.json) |
| Wall clearance, Pavo20 Pro | body stops 47.2 mm (median) beyond its own size on 5 cm voxel collision | 200 directions from 12 points; a record, not a gate, so it has no negative control (a8-clearance.json) |
| Stick-to-screen latency | display-limited: the test screen runs at 30 Hz (median 252 ms, blank-page floor 65 ms) | SendInput → Desktop Duplication marker, N = 220, lagFrames=2 control fires; camera measurement not done yet (a9-latency.json) |
superspl.at link or id ──► CDN: settings.json, lod-meta.json, scene.voxel.{json,bin}
│
┌──────────────┐ ┌───────────────┴──────┐ ┌──────────────────────┐
│ render │ │ sim-core (no DOM) │ │ input │
│ PlayCanvas │◄──│ 1 ms fixed step, 4 │◄──│ WebHID (EdgeTX) │
│ WebGPU, GPU │ │ motors, mixer, rate │ │ Gamepad API │
│ sort, LOD │ │ PID, BF/Actual/KISS/ │ │ touch sticks │
│ streaming │ │ Raceflight rates, │ │ calibration wizard, │
└──────────────┘ │ gravity, input log │ │ arm gate │
└──────────┬────────────┘ └──────────────────────┘
▼
┌──────────────────────┐ ┌──────────────────────┐
│ collision │ │ crash │
│ voxel octree (vendored│──►│ Rapier wreck, debris, │
│ MIT), swept spheres │ │ chase camera, respawn │
└──────────────────────┘ └──────────────────────┘
- Scenes: any public SuperSplat scene by link (
superspl.at/scene/<id>,superspl.at/s?id=<id>) or id, loaded by the visitor's browser from the public CDN — the server never talks to SuperSplat. Scenes without collision data fly as "no walls" and say so (b-fly-b6.json,b-fly-b7.json). - Flight model: own code, re-implemented from Betaflight's published formulas (no GPL code copied): motor lag, battery sag, drag, airmode, angle mode for beginners. Every constant carries its source. Checked in Node (
a4-physics.json): hover, full-stick rates, motor lag, airmode at zero throttle, PID step response, drag (terminal velocity). Battery sag is only recorded as information (no pass/fail check); angle mode is not tested yet. - Crashes: impact speed decides; the wreck tumbles with debris (Rapier), then the crash screen offers a respawn at the start or at the last safe point.
b-fly-b12.jsonrecorded one wall hit, 4 debris pieces and a respawn into free space; it does not record which of the two points was used. Crash effects stay under the WCAG limit of 3 flashes per second: at most 2 flashes per second measured over 50 crashes, whole frame and quarters (b-fly-b17.json). - Gravity: Earth, Moon, Mars, zero-g or your own value, with an honest "same motors" mode and a "keep thrust-to-weight" mode. Measured on the live site: disarmed drops on Earth (9.59 m/s² for 9.81) and the Moon (1.617 m/s² for 1.62) (
b-fly-b16.json). That check has no real negative control yet: its control is a fixed flag in the harness, not a run that could fail. Mars, zero-g, a custom value and the "keep thrust-to-weight" mode are not tested yet in the app (zero-g is checked only in the Node core: angular momentum is conserved,a4-physics.json). - Replays: only stick inputs are stored; the flight is recomputed bit-exactly (
b-fly-b15.json). - Walls for scans published without them: one click builds collision in the browser with SuperSplat's own tool and defaults (
@playcanvas/splat-transform, 5 cm voxels). On a 2.1 M-Gaussian scan it takes about 25 s and the result is byte-for-byte identical to the command-line tool; scans above 4 M Gaussians are refused with the numbers (c-bake.json, which has no negative control of its own; 0 penetrations in 50,000 passes on the baked walls, with a control,c-a5-baked.json). - Cinema mode: finest level of detail, automatic quality off, and a WebCodecs recorder (H.264 MP4) for the showcase scans, with the scene's credit burned into the picture (
d-cinema.json: a 6 s, 179-frame test recording; the credit strip was checked on one frame, not frame by frame). - Quality governor: when frames start missing the display's refresh, render scale and splat budget step down, and come back when the GPU has room again. Measured: render scale 1 → 0.5 under load and back to 1 (
d-governor.json); the splat-budget step is not measured yet. - Your Betaflight settings: paste
diffordiff allfrom the Betaflight CLI (4.3–4.5, 2025.12) and the simulated drone gets your rates, PID and throttle curve; anything the simulator does not use is listed, anything ambiguous is refused with the reason. Checked on the live site for rates and PID with a 4.5.1 and a 4.4.3 diff written in the firmware's exact print format but with made-up values (d-bf-diff-import.json); the throttle curve, 4.3 and 2025.12 are covered by unit tests only, and diffs from real flight controllers are not tested yet. - Trajectory export: CSV or JSON at 100 samples per second, straight from the flight model. The CSV was checked on a 12.6 s flight (1,256 rows): the 125 positions sampled every 0.1 s equal the flight recomputed from the input log exactly, and a one-row shift is caught (
d-trajectory-export.json); the JSON export is not tested yet.
Architecture decisions and their reasons: docs/decisions.md. What can go wrong: docs/warnings.md. Where every piece of code came from: docs/PROVENANCE.md.
| Browser | Input | Status |
|---|---|---|
| Chrome, desktop | EdgeTX radio (RadioMaster, Jumper, …) as USB joystick over WebHID | tested in Chrome with a simulated EdgeTX radio; real radio not tested yet; Edge not run yet |
| Chrome / Edge, desktop | Gamepad, DJI controller as a gamepad | not tested on real hardware yet |
| Desktop Chrome, touch emulation | Touch sticks | tested only in emulation (375×812, 1024×768): the arm button, both sticks at once and the pause button work. Holding a hover was not checked (the craft sat 0.3–3.3 m above its hover target), and the check's negative control is not conclusive yet. Safari, iPad and Android browsers not tested yet |
| Firefox 155 (Playwright build) | Gamepad or touch sticks | the scene renders with WebGPU and the app offers a gamepad or touch sticks (flying there not tested yet); radios need Chrome or Edge (no WebHID) |
Have an EdgeTX radio? A short test report is the most useful contribution right now — open an issue.
How can I fly an FPV drone through a 3D Gaussian Splatting scan? Open https://gsfpv.flyreelstudio.eu, pick one of the showcase scans or paste any public SuperSplat link, connect your radio, gamepad or touch sticks, arm and fly. No install, no account.
Is there a free FPV simulator that runs in the browser? Yes — ShramkoGSFPV is free and MIT-licensed. It needs WebGPU, so desktop Chrome or Edge works best.
Can I use my RadioMaster / EdgeTX radio in the browser? Yes: set the radio to USB Joystick mode, open the simulator in Chrome or Edge, and the calibration wizard maps sticks, inversion and the arm switch. It has been verified only in Chrome with a simulated EdgeTX radio (Edge not run yet); real-radio reports are welcome.
Do I crash when I hit a wall, or fly through it? You crash. The drone's body is swept against the scan's voxel collision every 1 ms physics step, so even at 34 m/s it does not slip through a 2 cm wall: 0 penetrations in 50,000 passes at a 2 cm wall and in 150,000 more on three real scans.
Is the flight physics like Betaflight? The rate curves match Betaflight 4.5.1's own code with 0 error at 1880 points when both are computed in double precision (a float32 build differs by at most 6.3e-4 °/s); the PID and filters follow Betaflight's structure and default gains. It is a simulator, not a certified replica.
Can my AI agent set this up? Yes — see below.
Paste into Claude Code, Codex, Cursor or any coding agent:
Set up ShramkoGSFPV for me by following
https://github.com/AndriiShramko/ShramkoGSFPV/blob/main/AGENT_SETUP.md exactly.
Install, run the checks, start the simulator locally and open it in Chrome.
Do not change any system settings; ask me only if a step needs my password or a physical action.
Rules for agents working in this repository: AGENTS.md. Summary for language models: llms.txt.
git clone -c core.autocrlf=false https://github.com/AndriiShramko/ShramkoGSFPV.git
cd ShramkoGSFPV && corepack enable && pnpm install --frozen-lockfile
pnpm -r typecheck && pnpm vitest run
pnpm --filter @gsfpv/fly dev # http://localhost:5190/fly/?scene=39e63ce9Full steps, update and removal: AGENT_SETUP.md.
Most useful now: real-radio test reports (EdgeTX version, radio model, RF on/off), measurements on other GPUs and high-refresh displays, and flight-model review from pilots. See CONTRIBUTING.md and SECURITY.md.
Andrii Shramko — FPV pilot and 3D/4D Gaussian Splatting specialist, author and maintainer of ShramkoGSFPV.
I am passionate about this idea, always open to interesting people and ready to help teams bring incredible projects to life. If you are an FPV pilot, scanning studio, drone maker, simulator developer, 3DGS platform, researcher or investor — let's talk about partnership, integration, licensing, sponsorship or funding. I am ready to assemble and lead a team around it.
- LinkedIn: https://www.linkedin.com/in/andrii-shramko/
- Book a call: https://calendar.app.google/Ff729HqGk4RpzPNDA
- Email: zmei116@gmail.com
- GitHub: https://github.com/AndriiShramko
- Contact form: https://gsfpv.flyreelstudio.eu/en/#contact
Code is released under the MIT License © 2026 Andrii Shramko. If you use this project, keep the copyright notice and please credit Andrii Shramko with a link to this repository; citation metadata is in CITATION.cff.
Built on the MIT-licensed PlayCanvas Engine and SuperSplat viewer (collision code vendored, see NOTICE), Rapier (Apache-2.0) for the wreck, and formulas read from Betaflight 4.5.1. Thanks to SplatFPV, an earlier browser experiment that proved the idea works. Scenes remain the property of their authors and are shown with their license and attribution; authors can ask for removal through the "Report this scene" link in the simulator.
Product names (SuperSplat, PlayCanvas, Betaflight, EdgeTX, RadioMaster, BetaFPV, Liftoff, VelociDrone, DJI) belong to their owners and are used only to describe compatibility. This project is not affiliated with or endorsed by any of them.