ai4ce/DiscoNet

DiscoNet (NeurIPS 2021) is a seminal research project in the niche of Multi-Agent Perception, specifically for connected autonomous vehicles (V2X). Its primary innovation is using Teacher-Student Knowledge Distillation to train a 'student' model (which only has access to constrained, compressed wireless data from other cars) to mimic a 'teacher' model (which has access to the full, uncompressed sensor data of all agents). While highly influential in the academic sub-field, the project scores a 4 for defensibility because it is a static reference implementation rather than a living framework. The repository has 151 stars and 17 forks but zero current velocity, indicating it has moved from an active project to a historical benchmark. It faces high platform domination risk from automotive OEMs and Tier 1 suppliers (like Bosch, Wayve, or Cruise) who would implement these algorithms within proprietary, vertically integrated stacks. Furthermore, newer architectures like V2X-ViT (Vision Transformers) have largely superseded the GNN-based approach used here. The 'moat' is purely academic citation gravity rather than software lock-in.