Growing from Exploration: A self-exploring framework for robots based on foundation models

Shoujie Li,Ran Yu,Tong Wu,JunWen Zhong,Xiao-Ping Zhang,Wenbo Ding

2024-01-24

Abstract:Intelligent robot is the ultimate goal in the robotics field. Existing works leverage learning-based or optimization-based methods to accomplish human-defined tasks. However, the challenge of enabling robots to explore various environments autonomously remains unresolved. In this work, we propose a framework named GExp, which enables robots to explore and learn autonomously without human intervention. To achieve this goal, we devise modules including self-exploration, knowledge-base-building, and close-loop feedback based on foundation models. Inspired by the way that infants interact with the world, GExp encourages robots to understand and explore the environment with a series of self-generated tasks. During the process of exploration, the robot will acquire skills from beneficial experiences that are useful in the future. GExp provides robots with the ability to solve complex tasks through self-exploration. GExp work is independent of prior interactive knowledge and human intervention, allowing it to adapt directly to different scenarios, unlike previous studies that provided in-context examples as few-shot learning. In addition, we propose a workflow of deploying the real-world robot system with self-learned skills as an embodied assistant.

Robotics,Artificial Intelligence

What problem does this paper attempt to address?

The core problem that this paper attempts to solve is to enable robots to autonomously explore various environments without human intervention and continuously learn new skills in the process. Existing work mainly relies on learning - or optimization - based methods to complete human - defined tasks, but how to make robots explore autonomously in multiple environments remains a challenge. For this reason, the paper proposes a new framework named GExp, which is based on foundation models (such as large - language models and vision - language models) and aims to enable robots to understand the environment through self - generated tasks, learn skills from beneficial experiences, and ultimately solve complex tasks. GExp does not rely on any prior interaction knowledge or human intervention, which enables it to directly adapt to different scenarios without providing context examples for few - shot learning.

Specifically, the main contributions of the GExp framework include:
- Proposing a robot autonomous exploration framework that does not require specific human - defined tasks or prior knowledge about the environment, and its main function is to promote the continuous independent exploration of robots.
- By leveraging successful experiences, enabling robots to learn during the exploration process and generate zero - shot general skills, which not only enhance the robots' ability to solve complex tasks but also expand their capacity boundaries.
- Creating a self - verification module using pre - trained vision - language models to analyze and determine the success of task execution, achieving "backtracking control" by evaluating the pre - conditions of each task step, ensuring that actions are consistent with the overall task goals, and improving the precision and effectiveness of robots.
- Designing a series of experiments to effectively evaluate the robots' exploration and self - learning abilities and verify the feasibility and effectiveness of the proposed framework.

Through these contributions, the GExp framework aims to promote the development of robotics technology and make it closer to a truly autonomous intelligent system.