Nipping the Drift in the Bud: Retrospective Rectification for Robust Vision-Language Navigation

TL;DR

This work addresses exposure bias and Instruction-State Misalignment in Vision-Language Navigation by introducing BudVLN, an online framework that learns from on-policy rollouts. It couples GRPO-based optimality seeking for proficient samples with Retrospective Rectification (SFT) to prevent errors by re-anchoring supervision to valid historical states, ensuring semantic alignment with natural language instructions. The approach yields state-of-the-art results on R2R-CE and RxR-CE while reducing training cost relative to DAgger-based methods, validating the effectiveness of on-policy, semantically consistent supervision in long-horizon VLN tasks. Overall, BudVLN advances robust embodied navigation by tightly coupling exploration, on-policy supervision, and misalignment-free corrective guidance.

Abstract

Vision-Language Navigation (VLN) requires embodied agents to interpret natural language instructions and navigate through complex continuous 3D environments. However, the dominant imitation learning paradigm suffers from exposure bias, where minor deviations during inference lead to compounding errors. While DAgger-style approaches attempt to mitigate this by correcting error states, we identify a critical limitation: Instruction-State Misalignment. Forcing an agent to learn recovery actions from off-track states often creates supervision signals that semantically conflict with the original instruction. In response to these challenges, we introduce BudVLN, an online framework that learns from on-policy rollouts by constructing supervision to match the current state distribution. BudVLN performs retrospective rectification via counterfactual re-anchoring and decision-conditioned supervision synthesis, using a geodesic oracle to synthesize corrective trajectories that originate from valid historical states, ensuring semantic consistency. Experiments on the standard R2R-CE and RxR-CE benchmarks demonstrate that BudVLN consistently mitigates distribution shift and achieves state-of-the-art performance in both Success Rate and SPL.

Figures (3)

• Figure 1: Illustration of instruction-state misalignment. The yellow line in (a) depicts a static expert demonstration. When the online rollout deviates due to navigational uncertainty (red line in (b)), standard DAgger forces a recovery from the error state. However, the required backward-correcting actions (e.g., turning back (blue line in (c))) fail to establish a semantic connection with the instruction walk straight, leading to further grounding confusion. In contrast, our BudVLN employs retrospective rectification (green line in (d)): it re-anchors to the latest progressed point on the reference path to synthesize a forward-looking demonstration that remains strictly aligned with the natural language instruction.
• Figure 2: Overview of the BudVLN training framework. The framework employs an Adaptive Mutual Exclusion Strategy to harmonize exploration and supervision. For a given instruction, a greedy probe first evaluates the agent's proficiency. When proficient, the framework routes to the optimality seeking pathway, employing GRPO to reinforce high-SPL behaviors via diverse sampling. Conversely, upon failure, it triggers the rectification pathway, which reverts to a valid historical state to synthesize alignment-preserving supervision. Gradients are back-propagated exclusively from the active pathway to ensure stable policy updates.
• Figure 3: Qualitative comparison between the Baseline and BudVLN. The top row illustrates the Baseline agent failing to ground the instruction, leading to a deviation from the correct path and eventually getting stuck. In contrast, the bottom row demonstrates that BudVLN, trained with our Retrospective Rectification and GRPO, successfully navigates the complex environment. Despite visual similarities in the initial steps, BudVLN exhibits superior robustness, effectively avoiding the failure mode that trapped the baseline.