AIRSHIP: Empowering General-Purpose Intelligent Robots through Open-Source Embodied AI
Year: 2025
Type: preprint
Abstract: Embodied AI is transforming robotics from specialized systems to general-purpose agents capable of adapting to dynamic environments, unlocking new applications in industries and daily life. This paper introduces AIRSHIP, an open-source framework that integrates large language models (LLMs) with modular robotic components for high-level reasoning and low-level control in navigation and grasping tasks. A robust re-planning mechanism leverages LLM feedback to handle failures through re-execution, refinement, and exploration strategies, improving success rates from 50% to 90% for single directive tasks and from 6% to 70% for sequential tasks in evaluations. This mechanism dynamically interacts with semantic perception, navigation, and grasping modules. Under the modular framework, these components can be easily expanded or replaced, enabling AIRSHIP to adapt to other platforms and facilitate rapid iteration. Finally, we discuss the roadmap toward more general and adaptable robotic agents, suggesting possible directions to improve robustness, flexibility, and cross-task generalization. The source code of this work has been made publicly available at https://github.com/airs-cuhk/airship.
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4416888484