AIAny
Icon for item

OmniEcho: Audio-Visual Spatial Understanding for Omni-Modal Embodied Agents

Provides a unified benchmark and a spatially aware omni-modal model for spatial audio-visual reasoning in embodied agents — OmniEchoBench (197 scenes, 2,972 QA pairs; 900 FOA-based navigation tasks) plus OmniEcho, which adds an FOA spatial encoder to a pretrained semantic audio pathway to enable spatial QA and sound-guided navigation.

Introduction

Most embodied agents lack a systematic way to evaluate and learn spatial hearing: humans effortlessly localize and combine audio direction with vision, but existing omni-modal models miss a dedicated spatial-audio channel and there is no unified real-world benchmark. This work supplies both a large real-world evaluation suite and a model architecture that explicitly separates semantic and spatial audio pathways, making spatial audio a practical signal for embodied reasoning and navigation.

Key Findings
  • OmniEchoBench: a real-world unified benchmark comprising spatial audio-visual perception and sound-guided navigation. It contains 197 spatial scenes with 2,972 QA pairs for perception and 30 real-world environments yielding 900 navigation tasks with dense first-order ambisonics (FOA) recordings.
  • Scalable data pipeline: a controllable FOA rendering pipeline that preserves geometric consistency between sound sources, visuals, and agent trajectories for both perception QA and navigation trajectory synthesis.
  • Model design: OmniEcho augments a pretrained semantic audio pathway with a dedicated FOA spatial encoder and uses a three-stage alignment training to jointly capture auditory semantics and spatial cues while retaining semantic audio pretraining benefits.
  • Empirical results: OmniEcho attains state-of-the-art performance on the spatial audio-visual perception tasks and achieves sound-guided navigation performance approaching some traditional text-guided VLN baselines (example metrics: OmniEcho reported SR 16.2% and SPL 11.5% on OmniEchoBench-Nav), while highlighting persistent challenges in fine-grained localization, distance estimation, path efficiency, and reliable stopping.
Who it's for and tradeoffs

Great fit if you research multimodal embodied perception or want a real-world benchmark and baseline model to study spatial audio integration in navigation and QA. The dataset and model are valuable when you need FOA spatial audio, geometric consistency between audio and visuals, and reproducible navigation tasks. Look elsewhere if your focus is purely simulated binaural audio, large-scale web-scale audio-only tasks, or if you need out-of-the-box high-efficiency navigation systems — the work prioritizes spatial-audio understanding and diagnostic evaluation over immediate state-of-the-art VLN path efficiency.

Information

  • Websitearxiv.org
  • OrganizationsSchool of Intelligence Science and Technology, Peking University, State Key Laboratory of General Artificial Intelligence, Peking University, Alibaba Token Hub, Alibaba Group, Tsinghua University
  • AuthorsRuixun Liu, Yuxuan Wang, Jiacheng Xie, Yuhuan You, Donghua Cai, Junming Lin, Xiong-Hui Chen, Zhifang Guo, Yunfei Chu, Qize Yang …
  • Published date2026/09/20

More Items

Alternates a Planner (issues sub-queries) and a Synthesizer (integrates retrieved evidence into a persistent summary) to tackle long-horizon deep-search; introduces Role‑Decoupled Policy Optimization (RDPO) for role-specific RL credit assignment and shows strong results (IterSynth-8B reaches 50.7% on five benchmarks).

A mobile-planning agent paper that develops a closed-loop AI-for-AI lifecycle to generate data, train a planner model, and co-evolve a runtime Harness for multi-app task execution. Demonstrates top performance on MobilePA-Bench with improved tool use, memory, skill coordination and low estimated per-task output cost.

Edits an LLM agent's internal task state and action continuations to remove unsupported assumptions and noisy plans, improving downstream decisions instead of reconstructing high-entropy tool outputs. Combines Action Judge, State Revision, and EditAct and shows consistent gains across benchmarks and agent backbones.