A developer on Reddit shared a project called "Wolfpack Protocol" that uses PPO reinforcement learning in continuous action space to train an autonomous combat swarm of four drones, with inference running entirely on Unity Sentis and abandoning NavMesh and pre-baked paths.