The open arena where AI learns to fly.
Train a drone pilot, prove it against the world, and watch it fly.
Drones are moving into everyday life: delivering packages, inspecting bridges and power lines, searching for survivors after a disaster. The hardware is ready. The hard part is the AI that flies them, and today that AI is built behind closed doors, where no one can compare it or prove whose is best.
Swarm is the open arena that settles it. Anyone can train a drone pilot, submit it, and watch it compete on a public leaderboard across thousands of fresh worlds it has never seen. No private test sets, no memorizing, no shortcuts. The best pilot wins, out in the open, and its rewards are paid automatically by the Bittensor network (Subnet 124).
swarm_sar_compressed.mp4
Search and Rescue. The mission Swarm is built around: teaching a drone to find people.
5 missions · 1,100 fresh worlds every week · 60-second flights · one live leaderboard
Swarm is five separate competitions, each with its own champion and its own share of the rewards. Master one, or take on all five.
| Mission | What your drone has to do | Reward share |
|---|---|---|
| Interceptor | Chase down a fleeing drone and catch it before time runs out |  |
| Swarm Autopilot | Land a whole team of drones, fast and without collisions |  |
| Swarm Search and Rescue | Send a team of drones to sweep the area and find the victim together | |
| Autopilot | Cross the world, find the landing pad, and touch down clean |  |
| Search and Rescue | Find a lost person and hold a steady hover above them | |
Every mission has a full guide with everything a builder needs to start.
swarm_interceptor_720p.mp4
Air-to-air pursuit, from a real benchmark run: close the gap and catch a fleeing drone before the clock runs out.
Every mission runs the same simple loop:
| Depth Camera what's in front of it |
→ Your Model → | Flight Commands where to fly next |
→ Drone |
| Flight State position and speed |
Third-person view |
What the drone sees |
The drone sees the world through a single depth camera and knows its own position and speed. Fifty times a second, your model looks at that and decides where to fly next: a direction, a speed, a turn. No map, no GPS, no list of obstacles. Just like a real pilot, it has to read what is in front of it and react. What changes between the five missions is the goal: land, rescue, coordinate a team, or give chase.
None of these worlds exist until the benchmark builds them. Every week, every drone flies 1,100 brand-new ones, so nothing can be memorized.
City streets, buildings, intersections |
Ski Village snow-roofed streets, mountains |
Mountains peaks and valleys |
Warehouse indoor racks and cranes |
Forest dense trees, tight gaps |
Open terrain wide skies, no cover |
Normal |
 Autumn |
 Snow |
 Dead |
Every flight earns a score from 0 to 1, on three things:
| Success did it finish the mission? 45% |
Speed how fast, within the time limit? 45% |
Safety how clear of obstacles? 10% |
A drone's rank is its average across all 1,100 worlds, so steady skill beats a few lucky runs. (Interceptor is pure pursuit: half success, half speed.)
Rewards run on King of the Hill. Each mission pays its last five champions, not just the current one, and your share depends on how much you raised the bar when you won. Beat the record and you keep earning even after someone beats you, so a real breakthrough pays off for a long time. The full mechanics are in the King of the Hill guide.
Everything happens from the terminal, with no dashboards to wrangle. Train, package, and score your model with a handful of commands:
swarm doctor # check your setup
swarm model package --source my_model/ # package your agent
swarm benchmark --model submission.zip --workers 4 # score it locally
swarm report # see how it didThe same engine validators use runs on your own machine, so the score you see locally is the score that counts. Full reference in the CLI docs.
Important
One shot per drone, per mission. Once your model is scored, its slot is locked. Practice locally as much as you want; when you are ready, submit your best and it competes for good.
|
Build a drone pilot from scratch. The Miner Guide walks you through the whole path, from first install to a spot on the leaderboard. 
|
Help score the network on your own hardware. The Validator Guide covers setup, launch, and auto-updates. 
|
The pilots trained here don't stay in simulation.
Langostino is the open-source drone we built to prove it flies for real: 3D-printed, off-the-shelf parts, full build guide included. Anyone can make one.
Train in simulation → Compete on the leaderboard → Fly on real hardware
Distributed under the MIT License. See LICENSE for details.
