OmniStream — Continuous perception, reconstruction and action for embodied AI

Continuous perception is shifting toward persistent latent memory and interactive, camera-conditioned world models. LatentStream adds progressive internalization into fixed-length memory, while WorldReward provides action-aligned chunking and decomposed reward modeling for camera-conditioned worlds; these complement Think While Watching, Spatial-TTT, HSImul3R, SimRecon, M^3, Kinema4D, and WorldCam. Robustness to occlusion, tactile sensing, long-horizon memory, and sim-to-real transfer remain open.