SWARM (Subnet 124): Bittensor's Open Autonomous Drone Benchmark

Closed-door drone labs keep flight software behind an NDA and ask you to take their word for how well it flies. SWARM (Subnet 124) took the opposite bet, publishing the seeds, the scoring formula, and letting a model called SOTAPilot answer for itself. It landed a drone on an unseen platform 95.34% of the time, across six procedurally generated worlds, using nothing but a depth camera and its own flight state to find the way down.

The subnet does not hand out second chances either. You get exactly one submission per hotkey, the wallet address tied to your entry on the network, for life, and you have to beat the reigning champion's score by 0.015 in local testing before the network will even run the full benchmark on your model. There's no patch after launch and no resubmission if your training run comes up short.

SOTAPilot cleared the bar in public, and the trained model behind it, known in the field as a policy, now flies outside the simulator on an open-source quadcopter the team built to prove it. SWARM marked the milestone in its own July 2026 letter to stakeholders. The same letter announces a robotics lab opening in Andorra, a spot in the country's startup accelerator, and a push for government funding to build a mountain search-and-rescue drone, on the same terrain SOTAPilot already trains against.

Closed-Door Drone AI Finally Gets a Scoreboard

Delivery drones, inspection drones, and search-and-rescue drones keep shipping hardware, but the flight software behind them stays locked inside individual companies. You have no way to check whether one team's navigation policy beats another's, because no shared benchmark exists to test it. SWARM built one and put it on-chain, open for you to run yourself.

You provide a trained navigation model, built and tested on your own machine. SWARM provides the infrastructure to prove it works, a public benchmark, a fresh batch of worlds every week, and if your model ranks well, a share of the TAO that the subnet pays out.

Submit your model under your hotkey, and validators drop it into a physics simulator. It receives a 128×128 depth image and a state vector covering position, velocity, and orientation, fifty times a second, and answers with a five-dimensional velocity command. It has sixty seconds to fly to a landing platform in a world it has never seen before.

Your model has to clear six environment types before you know if it works, from dense city blocks and a twelve-meter warehouse stacked with racks and cranes to open mountain terrain, long-range open terrain, and a forest that shifts through four seasonal looks, from full canopy to bare winter branches.

Validators generate 1,000 unique seeds every epoch, refreshed every seven days. Each seed is a single starting number that regenerates the same world on demand, and SWARM publishes every one on Swarm124, so you never draw an easier map than the miner next to you.

Bittensor already runs subnets that touch physical-world data. NATIX crowdsources street-level video from drivers to improve maps, and Loosh is building a cognition layer meant to give robots memory and judgment.

Neither one scores a flight-control policy against a public leaderboard. SWARM is the subnet built specifically to put autonomous drone navigation through an open, adversarial benchmark.

One Shot, No Retries, and a 95.34% Score to Beat

Score your model on three weighted pieces:

• Success (45%): did the drone reach the landing platform
• Speed (45%): how fast, relative to the time limit
• Safety (10%): minimum clearance kept from obstacles during flight

Your rank comes from the average across all 1,000 seeds, which rewards you for surviving every run instead of getting lucky on a handful. Getting onto the leaderboard at all means beating the current champion's score by 0.015 first, in your own local testing, before the network runs the official benchmark on what you submitted.

You'd be one of many trying. SWARM's own count, posted when SOTAPilot shipped, shows over 400 submissions against 25 that beat the champion's score. Most models never clear their own local test, let alone the network's.

Clear it, and your hotkey locks to a single model, permanently, with no retraining once it's live. SOTAPilot, published in the subnet's own repository, cleared the same gate at a 95.34% success rate across all six environment types, reading the depth image and its own flight state alone, with no hand-built map and no GPS fix to fall back on.

The Model Leaves the Simulator

A benchmark score means little if the model never leaves the container it was scored in. SWARM built Langostino to close the gap, an open-source quadcopter running ROS2, Raspberry Pi, and INAV firmware, with a full parts list and 3D-printable components published alongside the flight code, ready for you to build.

Train in simulation, climb the public leaderboard, then load the same policy onto Langostino and watch it fly outdoors on real hardware, without a rewrite. Score well only in PyBullet physics, and your model never reaches Langostino at all.

Andorra Is Where the Thesis Gets Tested

SOTAPilot already made the leap. What comes after is less about the model and more about where SWARM puts it to work.

SWARM turned one year old around its July 2026 letter to stakeholders, and the team spent the year moving the subnet's ambitions off a leaderboard and into a country. A robotics lab in Andorra is in its final setup stages, funded in part by a €370,000 raise from EverBlueGreen, which the team says covers the lab, hiring, and the shift from demonstrations to repeatable deployments.

Andorra's government selected SWARM for Enlaira, its official startup acceleration program, as one of five companies chosen nationwide. The next target is a €200,000 to €300,000 grant from Andorra Business, earmarked for a fully autonomous drone system built for mountain search and rescue, the same kind of terrain SOTAPilot already trains against in the benchmark's mountain and ski-village worlds.

If the grant lands, your stake in SWARM stops being a bet on a leaderboard score and starts being a bet on a government contract.

SOTAPilot Was Built to Outgrow Its Own Benchmark

SWARM's plan pushes SOTAPilot's benchmark beyond single-drone navigation, toward search and rescue, interception intelligence, and coordinated multi-drone flight. If you're training a model on the subnet next, you're training it on those tasks, not as a research afterthought.

Coordinating a swarm on a shared mission is a different engineering problem than landing one drone on one platform. Your model has to account for other drones sharing the same airspace, beyond the terrain in front of the camera.

SWARM wants the benchmark to become the reference standard for autonomous drone piloting, open and reproducible, tied to hardware flying outside the simulator rather than another leaderboard screenshot.

Bittensor's First Real Flight Test

Closed drone labs ask you to trust the marketing copy. SWARM publishes the seeds, the scoring formula, and the champion's score for anyone to check, then dares a new model to beat it by 0.015 before the network even runs a benchmark on it.

SOTAPilot cleared the bar at 95.34% with no map and no GPS, and the same policy now flies on Langostino outside the simulator, with an Andorra acceleration slot already behind the next step.

A subnet tells you what it built in a whitepaper. SWARM shows you: a drone clearing a warehouse it has never seen and landing in under a minute, in public, on a seed open for anyone to rerun.

Disclaimer: This article is for informational purposes only and does not constitute financial, investment, or trading advice. The information provided should not be interpreted as an endorsement of any digital asset, security, or investment strategy. Readers should conduct their own research and consult with a licensed financial professional before making any investment decisions. The publisher and its contributors are not responsible for any losses that may arise from reliance on the information presented.