SparseWorld-TC: Trajectory-Conditioned Sparse Occupancy World Model

SparseWorld-TC addresses a critical bottleneck in autonomous driving world models: the representational loss associated with VAE-based discrete occupancy tokens (used in projects like OccWorld or Drive-WM). By predicting multi-frame occupancy directly from image features in an end-to-end fashion, it offers higher fidelity and better trajectory alignment. However, with 0 stars and 8 forks at 3 days old, it is currently in the 'academic interest' phase. The defensibility is limited (4) because while the architectural approach is sophisticated, it is a point-solution in a hyper-competitive field dominated by well-funded labs (Waymo, Tesla, NVIDIA). The 'no-VAE' approach is a logical evolutionary step rather than a permanent moat. Frontier labs like OpenAI are building general world models (Sora), but the specific constraints of 3D sparse occupancy for robotics remain a niche that allows this project some breathing room (Medium frontier risk). The primary threat is displacement by more integrated 'end-to-end driving' models (like UniAD or VAD) that incorporate planning and perception into a single unified transformer, potentially making standalone occupancy world models redundant within 18-24 months.