Early to Share, Late to Save Synchronisation-Driven Communication Gating in Bandwidth-Constrained Cooperative VLN

Jul 16, 2026

Most cooperative Vision-Language Navigation (VLN) work assumes unlimited communication, while real deployments face strict bandwidth limits that require information-efficient coordination. The paper introduces bandwidth-constrained cooperative VLN and a hindsight gating method: a lightweight supervised gate trained post-hoc from navigation failures to avoid REINFORCE’s high variance. Results show counterintuitive behavior—gates activate mainly in early steps, even when agents are confident—explained by recurrent hidden-state synchronization.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

Introduction

Related Work

Problem Formulation

Hindsight Gating

Experiments and Analysis

Implications and Limitations

arXiv :2607 .08504v 1 [ cs .MA] 9 Jul 2026 Early to Share, Late to Save: Synchronisation-Driven Communication Gating in Bandwidth-Constrained Cooperative VLN Arav Gupta⋆ , Nivedan Yakolli, and Avinash Gautam Birla Institute of Technology and Science, Pilani Campus, Pilani, Rajasthan 333031, India [f20231280@pilani.bits-pilani.ac.in](f20231280@pilani.bits-pilani.ac.in) [p20230032@pilani.bits-pilani.ac.in](p20230032@pilani.bits-pilani.ac.in) [avinash@pilani.bits-pilani.ac.in](avinash@pilani.bits-pilani.ac.in) Abstract. Most cooperative Vision-Language Navigation (VLN) methods assume unlimited communication, not considering real-world applications where bandwidth is restricted and information efficiency is critical. We introduce bandwidth-constrained cooperative VLN and propose hindsight gating: a lightweight supervised gate that labels communication-critical steps post-hoc from navigation failures, avoiding the high variance of REINFORCE. Contrary to the intuition that agents should communicate when uncertain, we observe a consistent counterintuitive pattern: trained gates fire predominantly in early episode steps and more often when agents are confident, across all budget levels (B ∈{1, 3 , 5}) . We explain this through recurrent hidden-state alignment: early communication injects grounded trajectory representations that persist and compound through subsequent Gated Recurrent Unit (GRU) updates, achieving +0 .072 cumulative alignment gain with B=3 transmissions, approaching unconstrained communication (+0 .078) at 260% greater alignment efficiency than random gating (+0 .020) and 320% greater efficiency than entropy-based gating (+0 .017) . Our results establish a new communication regime for bandwidth-limited embodied agents: synchronise representations early, navigate independently later. Our codebase is available at: [https://github.com/AravG13/bandwidth](https://github.com/AravG13/bandwidth) [-constrained-cooperative-vln](-constrained-cooperative-vln.)[.](-constrained-cooperative-vln.) Keywords: Vision-Language Navigation · Multi-Agent Communication · Bandwidth-Constrained Communication · Recurrent Neural Networks · Hindsight Supervision. 1 Introduction Vision-Language Navigation (VLN) requires an embodied agent to follow natural language instructions through photorealistic indoor environments [1] . Real-world deployments such as search-and-rescue, warehouse automation, and multi-drone ⋆ Corresponding author: [f20231280@pilani.bits-pilani.ac.in](f20231280@pilani.bits-pilani.ac.in) 2 A. Gupta et al. inspection naturally involve multiple coordinated robots sharing a building anda goal. Extending VLN to this cooperative setting introduces a fundamental engineering constraint: radio channels, mesh networks, and privacy-constrained systems all impose hard limits on how often agents can communicate. Yet existing cooperative navigation methods [10, 15, 16] assume agents share observations freely at every step. This makes the when of communication – which steps are worth a transmission under a tight budget – an open problem for language-guided navigation. The challenge is compounded by a training difficulty: prior methods [11, 14] learn communication gates with REINFORCE, which suffers from high variance because the causal link between a single gate decision at step t and episode success 15–20 steps later is long and noisy. We address both challenges. First, we propose hindsight gating: rather than learning which steps to communicate through trial-and-error policy gradients, we run agents without communication, observe where each agent failed while its partner succeeded, and use those observations as direct Binary Cross-Entropy (BCE) supervision for a lightweight gate. This converts a high-variance policy gradient problem into a stable supervised classification problem with zero reward variance. Studying what the trained gate learns reveals a surprising finding. Rather than firing when ag

FAQs

What problem does the paper address in cooperative VLN?

It addresses cooperative VLN under bandwidth-constrained communication, where agents cannot freely exchange information at every step due to practical channel and privacy limitations.

How does hindsight gating train the communication policy?

Agents run without communication; where one agent fails while the partner succeeds, those observations become binary cross-entropy supervision to train a lightweight gate, avoiding the high-variance REINFORCE signal.

Why does the trained gate communicate mostly in early episode steps?

The paper attributes this to hidden-state alignment: early communication injects grounded trajectory representations into GRU hidden states, which then propagate and compound through later recurrent updates even after communication stops.

Jul 16, 2026

Upload by Valentina