swarm-subnet/swarm

Leverages Bittensor's decentralized subnet architecture where validators generate procedurally-varied drone navigation tasks (5-15m goal distances, challenge levels 1-4) and evaluate miner-submitted RL policies in isolated Docker containers using PyBullet physics simulation. Miners compete by training custom policies in any framework (Stable Baselines 3, PyTorch, JAX) against synchronized challenge seeds, with rewards split 50% on task success and 50% on execution speed. Winner-take-all distribution incentivizes fastest, most reliable autonomous flight controllers for micro-drone deployment.