Interactive Navigation of Quadruped Robots in Challenging Environments using Large Language Models

Kangjie Zhou, Yao Mu, Pengying Wu, Han Gao, Chang Liu

Keywords: Interactive navigation, large language model, embodied agent, closed-loop task and motion planning

Abstract:

Robotic navigation in complex environments remains a critical research challenge.
Notably, quadrupedal navigation has made significant progress due to the terrain adaptivity and movement dexterity of quadruped robots.
However, traditional navigation tasks confine the robot to a predefined free space and focus on obstacle avoidance, limiting their applicability in more challenging environments, such as scenarios lacking feasible paths to the target.
We propose an interactive navigation approach that leverages agile quadrupedal movements to adapt to diverse terrains and interact with environments, changing the workspace to tackle challenging navigation tasks in open and complex environments.
We present a primitive tree for high-level task planning with large language models (LLMs), facilitating effective reasoning and task decomposition for long-horizon tasks.
The tree structure allows for dynamic node addition and pruning, enabling adaptive responses to new observations and enhancing both robustness and real-time performance during navigation.
For low-level motion planning, we adopt reinforcement learning to pre-train a skill library containing complex locomotion and interaction behaviors for task execution.
Furthermore, we introduce a cognition-based replanning method consisting of the advisor and arborist to react to real-time egocentric observations.
The proposed method has been validated in multiple simulated scenarios, demonstrating its effectiveness in diverse scenarios and real-time adaptivity in partially observable conditions.

Submission Number: 8