StepNav: Structured Trajectory Priors for Efficient and Multimodal Visual Navigation

TL;DR

StepNav tackles the challenge of generating reliable, uncertainty-aware trajectories for visual navigation under perceptual ambiguity. It combines a geometry-aware success probability field with a structured multimodal prior, extracted from temporally refined features, and refines them with a regularized conditional flow-matching model to enforce safety and smoothness. Key contributions include a biharmonic-regularized PDE for the success field $F$, a K-shortest-path based diverse prior with energy $E(\tau)$ to capture multiple corridors, and a Reg-CFM objective with $L_{smooth}$ and $L_{safe}$ to drive safe, dynamically-feasible trajectories using few integration steps. Real-world validation on indoor/outdoor benchmarks and robotic deployment demonstrates superior robustness, efficiency, and safety compared to state-of-the-art generative planners, enabling practical real-time navigation.

Abstract

Visual navigation is fundamental to autonomous systems, yet generating reliable trajectories in cluttered and uncertain environments remains a core challenge. Recent generative models promise end-to-end synthesis, but their reliance on unstructured noise priors often yields unsafe, inefficient, or unimodal plans that cannot meet real-time requirements. We propose StepNav, a novel framework that bridges this gap by introducing structured, multimodal trajectory priors derived from variational principles. StepNav first learns a geometry-aware success probability field to identify all feasible navigation corridors. These corridors are then used to construct an explicit, multi-modal mixture prior that initializes a conditional flow-matching process. This refinement is formulated as an optimal control problem with explicit smoothness and safety regularization. By replacing unstructured noise with physically-grounded candidates, StepNav generates safer and more efficient plans in significantly fewer steps. Experiments in both simulation and real-world benchmarks demonstrate consistent improvements in robustness, efficiency, and safety over state-of-the-art generative planners, advancing reliable trajectory generation for practical autonomous navigation. The code has been released at https://github.com/LuoXubo/StepNav.

Figures (7)

• Figure 1: (a) FlowNav failed to avoid obstacles when navigating towards the goal (an UAV). (b) Our StepNav estimates the success probability field to generate the prior trajectory that can accelerate the refinement and improve the success rate.
• Figure 2: Overview of StepNav. Given a sequence of input images and a goal image, we first extract temporal features via V-JEPA2 and refine them using our DIFP module. These refined features are used to predict a success probability field $F$. From this field, we estimate several prior trajectories and sample the prior trajectory according to the energy landscape $E(\tau)$. This prior is then refined into a smooth, feasible trajectory via Reg-CFM.
• Figure 3: Examples of StepNav in complex indoor and outdoor navigation tasks. The top row shows outdoor navigation in a simulated urban environment, where StepNav successfully navigates a multi-lane intersection with other agents. The bottom row illustrates indoor navigation in a cluttered room, demonstrating avoidance of static obstacles. The yellow trajectories are the estimated paths and the green dots represent the groundtruth waypoints.
• Figure 4: A qualitative visualization of StepNav's core mechanism at an ambiguous T-junction. The gray lines indicate the generated prior trajectories, while the yellow lines represent the final trajectories of StepNav and other baseline models. The ground truth trajectories are shown in green.
• Figure 5: Refinement strategies under identical structured prior initialization. StepNav converges faster and yields smoother trajectories compared to CFM, which starts from Gaussian noise.