EmergeNav: Structured Embodied Inference for Zero-Shot Vision-and-Language Navigation in Continuous Environments

Abstract

Zero-shot vision-and-language navigation in continuous environments (VLN-CE) remains challenging for modern vision-language models (VLMs). Although these models encode useful semantic priors, their open-ended reasoning does not directly translate into stable long-horizon embodied execution. We argue that the key bottleneck is not missing knowledge alone, but missing an execution structure for organizing instruction following, perceptual grounding, temporal progress, and stage verification. We propose EmergeNav, a zero-shot framework that formulates continuous VLN as structured embodied inference. EmergeNav combines a Plan--Solve--Transition hierarchy for stage-structured execution, GIPE for goal-conditioned perceptual extraction, contrastive dual-memory reasoning for progress grounding, and role-separated Dual-FOV sensing for time-aligned local control and boundary verification. On VLN-CE, EmergeNav achieves strong zero-shot performance using only open-source VLM backbones and no task-specific training, explicit maps, graph search, or waypoint predictors, reaching 30.00 SR with Qwen3-VL-8B and 37.00 SR with Qwen3-VL-32B. These results suggest that explicit execution structure is a key ingredient for turning VLM priors into stable embodied navigation behavior.

Figures (4)

• Figure 1: Overview of EmergeNav. Given an instruction, the agent first decomposes it into anchor-grounded subgoals and then performs stage-structured embodied inference through a Plan--Solve--Transition hierarchy. The Solve stage executes high-frequency local rollouts under the active subgoal using forward-centered triplet views, while the Transition stage performs low-frequency boundary verification using panoramic observations. GIPE provides task-aligned perceptual evidence for both stages, and dual memory maintains dense within-subgoal traces (STM) together with sparse verified progress anchors (LTM).
• Figure 2: Unified inference loop of EmergeNav. The agent alternates between high-frequency local solving and low-frequency boundary verification while maintaining explicit evidence and memory states.
• Figure 3: GIPE interface in EmergeNav. In the solve stage, GIPE extracts compact local evidence from the active subgoal, forward-centered triplet views, and memory context to support action generation. In the transition stage, GIPE extracts boundary-level evidence from the current and next subgoals, panoramic observations, long-term memory, and rollout summary to support stage verification and Continue/Switch decisions.
• Figure 4: Contrastive dual-memory reasoning in EmergeNav. STM stores dense within-subgoal front-view traces, while LTM stores sparse verified progress anchors. Progress is grounded by comparing the evolving STM sequence against the LTM anchor, enabling the agent to distinguish continued advancement, stall, and revisit patterns during transition auditing.