EA-Nav: Learning Safe Visual Navigation Policies with Embodiment Awareness

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

EA-Nav, an imitation-learning-based framework, addresses cross-embodiment navigation, a key challenge in embodied intelligence, by proposing a modular multi-stage design that explicitly incorporates embodiment geometry to reduce action ambiguity under the same visual observation. The system constructs a cross-embodiment navigation dataset from Internet videos for pretraining, introducing embodiment geometry as conditional tokens. During fine-tuning, it uses a multimodal information injection mechanism with a decoupled architecture, including a trajectory augmentation strategy to generate high-risk samples. These samples train separate spatial perception and risk-aware correction modules. Experimental results on datasets like i2Nav, InternUtopia, and real-world robots (TurtleBot, Unitree Go2) demonstrate improved navigation performance across various embodiment settings, including agent radii from 0.5 m to 2.5 m, confirming the effectiveness of integrating embodiment geometry.

Key takeaway

For Machine Learning Engineers developing navigation systems for diverse robot platforms, you should explicitly incorporate embodiment geometry into your policy learning. This approach, demonstrated by EA-Nav, significantly improves cross-embodiment generalization and safety by resolving action ambiguity. Consider using a two-stage imitation learning framework with risk trajectory augmentation to enhance spatial perception and trajectory correction, especially for real-world deployment.

Key insights

Embodiment geometry is crucial for resolving action ambiguity in cross-embodiment visual navigation policies.

Principles

• Embodiment geometry resolves visual action ambiguity.
• Decoupled modules enhance risk perception and correction.
• Augmenting high-risk trajectories improves safety learning.

Method

EA-Nav uses a two-stage IL framework: pretraining with Internet videos and embodiment geometry tokens, then fine-tuning with a decoupled architecture for spatial perception and risk-aware correction, supported by risk trajectory augmentation.

In practice

• Condition navigation models on agent geometry.
• Augment training data with synthesized high-risk trajectories.
• Decouple risk perception from action prediction.

Topics

• Embodied AI
• Visual Navigation
• Imitation Learning
• Embodiment Geometry
• Risk-Aware Navigation
• Trajectory Augmentation

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• EA-Nav: Learning Safe Visual Navigation Policies with Embodiment Awareness — Incorporating embodiment geometry into imitation learning improves cross-embodiment visual navigation safety and performance.
• NavVerse: Benchmarking Indoor-to-Outdoor Embodied Navigation in Continuous Robot Simulation — NavVerse benchmarks continuous indoor-to-outdoor robot navigation, revealing current agents struggle with cross-context adaptation and safety.
• X-Diffusion: Training Diffusion Policies on Cross-Embodiment Human Demonstrations — X-Diffusion selectively integrates noisy human demonstrations into robot policy training to overcome embodiment mismatches.
• Beyond Waypoints: A Trajectory-Centric Waypointing Paradigm for Vision-Language Navigation — Grounding waypoints in executable trajectories resolves planning-control inconsistencies and improves reachability in Vision-Language Navigation.
• Path-level Hindsight Instructions for Semantic Exploration in Vision-Language Navigation — Phi-Nav uses hindsight instructions to align exploratory trajectories with language, generating dense training signals for VLN agents with limited data.
• Embodied Active Learning under Limited Annotation and Navigation Budget for Object Detection — Embodied active learning uses spatial inconsistency to efficiently adapt object detectors in new environments with limited budgets.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.