UniNav: A Unified World-Action Diffusion Model for Visual Navigation

2026-08-04

AI Summary

This work addresses the longstanding trade-off in visual navigation between computational efficiency and future scene understanding, where existing methods either lack visual foresight or rely on time-consuming planning rollouts. To overcome these limitations, we propose UniNav, the first framework that unifies future visual prediction and path planning within a single diffusion model. By integrating geometry-aware camera embeddings and fusing visual and waypoint tokens, UniNav simultaneously performs world modeling and action prediction within a unified Transformer architecture. Trained solely on video data without requiring trajectory annotations, our approach drastically reduces reliance on labeled paths while enabling efficient inference. Experiments demonstrate that UniNav consistently outperforms state-of-the-art methods across multiple navigation benchmarks; its accelerated variant, UniNav-Fast, achieves inference in just 0.1 seconds per step with negligible accuracy degradation.

Abstract

Image-goal visual navigation is a fundamental capability for embodied agents. Existing navigation policies efficiently predict waypoint trajectories but lack visual foresight, while navigation world models can anticipate future observations but often require costly planning rollouts. We present UniNav, a unified world-action model that generates future visual observations and continuous waypoint trajectories through a single diffusion process. Given history frames and a goal image, UniNav jointly denoises visual and waypoint tokens within a single transformer, unifying future prediction and action generation in a shared framework. To improve spatial grounding, we incorporate geometry-aware camera tokens. We also train on both trajectory-labeled navigation data and video-only data, enabling the model to benefit from diverse videos without waypoint annotations. Based on this unified framework, we introduce two variants: UniNav-Full jointly predicts interpretable future observations and their corresponding trajectories, while UniNav-Fast removes future-image tokens at inference for efficient trajectory prediction. Experiments on navigation benchmarks show that UniNav outperforms the strongest baseline in ATE across all datasets. With one-step inference, UniNav-Fast achieves a latency of 0.1s without a substantial accuracy drop. Code will be released.

Problem

• visual navigation
• image-goal navigation
• world model
• action prediction
• future observation

Innovation

• unified world-action model
• diffusion-based navigation
• geometry-aware tokens
• trajectory-visual joint prediction
• annotation-free video training