> ## Content Index > Fetch the complete content index at: https://www.tao.media/llms.txt > Use this file to discover other available public pages before exploring further. # Swarm (SN124) Opens New Drone Interceptor Competition - URL: https://www.tao.media/swarm-sn124-opens-new-drone-interceptor-competition/ - Published: 2026-08-28T14:58:22.000Z - Updated: 2026-08-28T19:24:08.000Z - Description: Swarm opened a new Interceptor competition on Bittensor Subnet 124, built for sim-to-real transfer, with real drone hardware going to the winning miner. - Author: Antonio Verrico - Tags: Swarm [Swarm ](https://swarm124.com/?ref=tao.media)(SN124) is the Bittensor subnet building an open-source autopilot for drones, and it determines the best flight model by having anyone who wants to compete score against the same public benchmark. The subnet opened a new Interceptor competition, and the team says the winner takes delivery of real drone hardware rather than a leaderboard position alone. Interceptor is one of five challenge families on the [subnet](https://github.com/swarm-subnet/swarm?ref=tao.media), each with its own champion and its own slice of emissions. The new competition replaces a task the team says the original Interceptor already hit 100% on. The team describes that phase as a warm-up and says the new version is built for sim-to-real transfer, with its physical lab and its own drones recreated inside the simulator. > Today we launch our first open competition to train a drone interceptor with real hardware waiting for the winner > > Original Interceptor hit 100%๐Ÿ. Warm-up over. > > Built for sim-to-real transfer. We recreated our lab and drones in simulation. > > Weโ€™re going for the home run [https://t.co/smnkc5UZMM](https://t.co/smnkc5UZMM?ref=tao.media) [pic.twitter.com/Wu5kGNuyIh](https://t.co/Wu5kGNuyIh?ref=tao.media) > > โ€” Swarm (@SwarmSubnet) [August 24, 2026](https://x.com/SwarmSubnet/status/2091840841467810182?ref%5Fsrc=twsrc%5Etfw&ref=tao.media) ### What Miners Have to Build Interceptor is air-to-air pursuit. Your drone lifts off 60 to 100 meters from an airborne [target](https://github.com/swarm-subnet/swarm/blob/main/docs/families/interceptor.md?ref=tao.media) and has 60 seconds to catch it. The target cruises until you close within 12 meters, then flees at 4.5 m/s with a sideways jink while your own speed caps at 6.0\. Only physical contact or closing inside 0.15 meters counts as a catch. Scoring splits half for the catch and half for the clock, with no safety term, because closing distance aggressively is the point of this challenge. Hit an obstacle, tip past 75 degrees, or run out of time, and the run scores 0.01. Your drone flies on a single forward depth camera that resolves out to 100 meters, four times the range the other Swarm families work with. A hint point lands somewhere within 10 to 40 meters of the target and refreshes every two seconds, so the final stretch of the chase runs on the camera alone. ### Why the Hardware Prize Matters Swarm builds its models for aircrafts it makes itself. Langostino, the 3D-printed drone the team open-sourced under MIT, runs ROS2 on a Raspberry Pi with INAV firmware, ships a full bill of materials, and holds position outdoors on published [video](https://github.com/swarm-subnet/Langostino?ref=tao.media) using a model the subnet trained. Recreating the lab and the airframes in simulation narrows the gap a winning policy has to cross on its first real flight. A model tuned against a generic simulated quadrotor scores well but flies badly, because the mass, the thrust curve, and the camera angle all differ from the machine it eventually lands on. ### What's at Stake on Emissions Interceptor draws 30% of Subnet 124's emissions, double what Autopilot or Search and Rescue receive, which is the largest slice on the subnet. Swarm's King of the Hill [rules](https://github.com/swarm-subnet/swarm/blob/main/docs/king%5Fof%5Fthe%5Fhill.md?ref=tao.media) pay the last five champions of each family, tapering 30% per step down the ladder, so a real improvement keeps earning through four more dethronings. Seven days without a new champion and the whole slice goes to the burn address until someone takes the throne. Every submission runs the full 1,100 seeds with screening disabled. Each hotkey carries one model into one family, and the slot locks the moment the score posts, so competing again means registering a fresh hotkey. The outgoing Interceptor champion was measured at 0.9949, catching the target in 1,098 of 1,100 worlds. That's the number a challenger has to beat, and the improvement floor sits between 0.015 and 0.005 depending on how close the leader is to a perfect score. The competition is open now. The code, the benchmark, and the full family spec are public on the Swarm [repo](https://github.com/swarm-subnet?ref=tao.media).