UniNav: Unified Diffusion Model for Visual Navigation and Foresight

Embodied agents require robust visual navigation capabilities, typically relying on policies that predict waypoints or world models that anticipate future observations. Existing approaches often separate these functions, leading to limitations in visual foresight or requiring computationally expensive planning. UniNav introduces a unified world-action diffusion model that addresses these challenges.This model generates both future visual observations and continuous waypoint trajectories through a single diffusion process. It uses a transformer to jointly denoise visual and waypoint tokens, integrating future prediction and action generation. By incorporating geometry-aware camera tokens and training on diverse data (both trajectory-labeled and video-only), UniNav enhances spatial grounding and leverages broader datasets. Experiments show UniNav surpasses strong baselines in navigation benchmarks, with a fast variant achieving low latency without significant accuracy loss.