Research Analyzer
← Back ICRA 2026

AINav: Large Language Model-Based Adaptive Interactive Navigation

Kangjie Zhou, Yao Mu, Haoyang Song, Yi Zeng, Pengying Wu, Han Gao, Chang Liu

PDF

AI summary

Key figure (auto-extracted from paper)
AINav enables quadruped robots to proactively manipulate obstacles and dynamically replan via LLMs to reach previously unreachable goals in complex environments.
Interactive navigation Large language models Task and motion planning Quadruped robots Adaptive replanning Reinforcement learning

Problem

Traditional navigation assumes fixed free space and fails when no viable path exists, while existing interactive methods lack real-world applicability or cannot strategically utilize movable obstacles in partially observable settings.

Approach

The system combines an LLM-driven primitive skill tree for task decomposition, a reinforcement learning-pretrained skill library for robust motion execution, and an adaptive replanning mechanism that triggers and adjusts plans based on egocentric observations.

Key results

  • LLM-driven primitive skill tree enables robust task decomposition and rapid plan adaptation
  • RL-pretrained skill library provides versatile locomotion and interaction capabilities
  • Adaptive replanning mechanism allows proactive and reactive adjustments without global scene knowledge
  • Validated in simulations and real-world experiments demonstrating effectiveness in diverse challenging scenarios

Why it matters

It extends robotic navigation beyond passive avoidance, enabling active environmental manipulation for disaster rescue, warehouse logistics, and other cluttered, dynamic applications.

Abstract

Robotic navigation in complex environments re- mains a critical research challenge. Traditional navigation meth- ods focus on optimal trajectory generation within fixed free workspace, therefore struggling in environments lacking viable paths to the goal, such as disaster zones or cluttered ware- houses. To address this problem, we propose AINav, an adap- tive interactive navigation approach that proactively interacts with environments to create feasible paths to achieve originally unreachable goals. Specifically, we present a primitive skill tree for task planning with large language models (LLMs), facilitating effective reasoning to determine interaction objects and sequences. To ensure robust subtask execution, we adopt reinforcement learning to pre-train a comprehensive skill library containing versatile locomotion and interaction behaviors for motion planning. Furthermore, we introduce an adaptive replan- ning approach featuring two LLM-based modules: an advisor serving as a flexible replanning trigger and an arborist for autonomous plan adjustment. Integrated with the tree structure, the replanning mechanism allows for convenient node addition and pruning, enabling rapid plan adaptation in a priori un- known environments. Comprehensive simulations and experi- ments have demonstrated AINav’s effectiveness and adaptivity in diverse scenarios. The supplementary video is available at: https://youtu.be/CjXm5KFx9AI.

Index terms

Task and Motion Planning Reinforcement Learning Legged Robots

Related papers