Swarm
SN124Autonomous drone flight network coordinating swarms of drones with AI
An open benchmark where AI learns to fly drones through 1,000 procedurally generated 3D worlds, using nothing but a depth camera and raw flight state. No maps, no shortcuts, just raw skill.
// The open benchmark for autonomous drone flight.
Swarm is a Bittensor subnet that turns autonomous drone navigation into an open competition. Miners train neural networks to fly drones through simulated 3D environments, while validators score their performance based on how quickly and accurately they reach a landing platform.
The simple version: Think of it as a robotics benchmark, like how ImageNet standardized object detection, but for drone flight. You give the AI a depth camera image and a flight state, and it must figure out how to navigate to a goal, no prior knowledge of the environment allowed.
Centralized equivalent: There is no direct open equivalent. Proprietary drone flight algorithms are typically locked behind NDAs and corporate R and D departments, making fair comparison impossible.
How it works:
- Miners submit pre-trained reinforcement learning policies (any framework, Stable Baselines 3, custom, whatever works) that must navigate drones through unseen environments using only a 128x128 depth image and a state vector.
- Validators generate random MapTasks with start-to-goal pairs, evaluate miner policies headless in a PyBullet physics simulator, and score based on success (50% weight) and time efficiency (50% weight).
- The problem it solves: Autonomous drone algorithms are developed behind closed doors. There is no standardized way to compare flight policies, which slows innovation and limits access to advanced navigation intelligence.
- The opportunity: Delivery drones, inspection bots, search and rescue, autonomous flight is a multi-billion dollar industry. Open benchmarks accelerate the entire field by giving researchers and companies a shared evaluation standard.
- The Bittensor advantage: Decentralized evaluation means no single company controls the benchmark. The incentive structure drives miners to continuously improve their models, creating a living leaderboard that evolves with the best available flight AI.
- Traction signals: The benchmark recently shipped version 4.0.0 with the SOTApilot upgrade, enhancing autonomous flight intelligence with faster iteration cycles. The subnet has an active X account (@SwarmSubnet), a public leaderboard at swarm124.com/benchmark, and a growing community on Discord.
Category: Reinforcement Learning | Centralized Competitor: No direct equivalent (proprietary systems only)
Swarm occupies a unique niche in Bittensor, the intersection of robotics, simulation, and open-source AI benchmarking. While other subnets focus on text, images, or data, Swarm tackles physical intelligence, teaching neural networks to navigate real-world 3D spaces.
Mechanism:
Validators create synthetic environments with procedurally generated terrain, including cities, mountains, warehouses, forests, and open fields. Each environment presents random start-to-goal pairs at radial distances of 10 to 30 meters. The AI gets a depth camera image and flight state, then outputs velocity commands to fly the drone. It has 60 seconds to land on the platform.
The scoring is balanced between success (did it land?) and efficiency (how fast?). This dual metric prevents miners from gaming the system with slow but reliable policies, and rewards both skill and speed.
The v4.0 release introduced containerized evaluation, which eliminates data leaks and ensures every model is tested on truly unseen environments, preventing memorization as a strategy.
Price is up 21.7% over the last 7 days and 15.8% over 30 days, suggesting growing interest in the subnet. At 0.00835 TAO with a market cap around 27,487 TAO, it remains accessible for new participants.
- Low participant count: Only 1 active miner reported on-chain, which limits competitive pressure and could affect benchmark quality if participation does not grow.
- Competition: Other simulation-based subnets and external benchmarks (AirSim, Flightmare) could pull attention, though Swarms Bittensor-native incentive model is a differentiator.
- Execution: RL-based drone navigation is notoriously difficult. Scaling from simulation to real-world flight remains a long-term challenge for the entire field, not just this subnet.
IntoTAO, out.
