Wolfpack Protocol
A Unity-based drone swarm AI project that uses PPO reinforcement learning to train four drones as an autonomous combat swarm, performing inference purely through neural networks via Unity Sentis without reliance on NavMesh or pre-baked paths.
Timeline (last 90 days)
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.
In this project, each drone casts 38 rays into 360-degree space to form "spatial whiskers," simultaneously detecting dynamic obstacles and teammate positions, with the output fed directly as observation input to the policy network.
The developer emphasizes that the project's policy network achieves "emergent obstacle avoidance without retraining," meaning the policy learns a general response to ray-distance patterns rather than fixed paths for specific maps.
In this project, each drone's flight decisions are handled by the neural network, while formation tactics are managed by a separate C# coordinator, the "Tactical Swarm Director," which evaluates line-of-sight raycast results and dot-product calculations of player blind spots and dynamically rotates attackers for suicide dives.
All Facts (4)
In this project, each drone casts 38 rays into 360-degree space to form "spatial whiskers," simultaneously detecting dynamic obstacles and teammate positions, with the output fed directly as observation input to the policy network.
50%UnverifiedThe developer emphasizes that the project's policy network achieves "emergent obstacle avoidance without retraining," meaning the policy learns a general response to ray-distance patterns rather than fixed paths for specific maps.
50%UnverifiedIn this project, each drone's flight decisions are handled by the neural network, while formation tactics are managed by a separate C# coordinator, the "Tactical Swarm Director," which evaluates line-of-sight raycast results and dot-product calculations of player blind spots and dynamically rotates attackers for suicide dives.
50%UnverifiedA 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.
50%