HongkunQLiu/InMoE

InMoE combines graph neural networks with mixture-of-experts routing for trajectory prediction—a technically sound novel combination, but the project shows zero adoption signals (0 stars, 0 forks, 10 days old, no velocity). The README lacks quantitative results, comparison tables, or reproduction details that would signal maturity. No evidence of community engagement, issue tracking, or documentation beyond basic description. The approach targets a well-studied problem (pedestrian/vehicle trajectory prediction) where multiple established baselines exist (Social-GAN, trajectron++, etc.). Frontier labs (Google, Waymo, Tesla) have dedicated autonomous driving teams that likely own similar architectures internally; they could trivially integrate this if valuable, but the niche positioning (specific to interaction modeling) provides some insulation. The technical contribution—combining interaction graphs with MoE gating—is credible but incremental within the trajectory prediction literature. Without early traction, user base, or academic publication signal, this remains a fresh research prototype with no defensibility. High risk of obsolescence if: (1) a frontier lab releases a competitive trajectory prediction model, or (2) a better-positioned OSS project with similar ideas gains adoption first. Medium frontier risk because this is a narrow domain (trajectory prediction) that labs build as a component, not a platform.