Yueguang Ge
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Yueguang Ge has a total of 14 co-authors. Here are the top co-authors by number of shared publications:
- Shaolin Zhang
State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China
6 shared publications Haitao Wang
State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China
5 shared publicationsShuo Wang
3 shared publications- Shuo Wang
Chinese Academy of Sciences
3 shared publications Wenkai Chang
2 shared publicationsDayong Wen
1 shared publication
Latest Publications
- 21 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Robotics and Automation Letters
- Vol. 10 (6) , 5951-5958
Double-Feedback: Enhancing Large Language Models Reasoning in Robotic Tasks by Knowledge Graphs
Large language models (LLMs) have demonstrated remarkable reasoning capabilities. However, in real-world robotic tasks, LLMs face grounding issues and lack precise feedback, resulting in the generated solutions deviating from the actual situation. In this paper, we propose Double-Feedback, a method that enhances LLMs reasoning by Knowledge graphs (KGs). The KGs play three key roles in Double-Feedback: prompting the LLMs to generate solutions, representing the task scenes, and verifying the solutions to provide feedback. We design structured knowledge prompts that convey the task knowledge background, example solutions, revision principles, and robotic tasks to the LLMs. We also introduce the distributed representation to quantify the task scene with interpretability. Based on the structured knowledge prompts and the distributed representation, we employ the KGs to evaluate the feasibility of each step before execution and verify the effects of the solutions after completing the tasks. The LLMs can adjust and replan the solutions based on the feedback from the KGs. Extensive experiments demonstrate that Double-Feedback outperforms prior works in the ALFRED benchmark. In addition, ablation studies show that Double-Feedback guides LLMs in generating solutions aligned with robotic tasks in the real world.
- 21 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Robotics and Automation Letters
- Vol. 10 (6) , 5951-5958
Double-Feedback: Enhancing Large Language Models Reasoning in Robotic Tasks by Knowledge Graphs
Large language models (LLMs) have demonstrated remarkable reasoning capabilities. However, in real-world robotic tasks, LLMs face grounding issues and lack precise feedback, resulting in the generated solutions deviating from the actual situation. In this paper, we propose Double-Feedback, a method that enhances LLMs reasoning by Knowledge graphs (KGs). The KGs play three key roles in Double-Feedback: prompting the LLMs to generate solutions, representing the task scenes, and verifying the solutions to provide feedback. We design structured knowledge prompts that convey the task knowledge background, example solutions, revision principles, and robotic tasks to the LLMs. We also introduce the distributed representation to quantify the task scene with interpretability. Based on the structured knowledge prompts and the distributed representation, we employ the KGs to evaluate the feasibility of each step before execution and verify the effects of the solutions after completing the tasks. The LLMs can adjust and replan the solutions based on the feedback from the KGs. Extensive experiments demonstrate that Double-Feedback outperforms prior works in the ALFRED benchmark. In addition, ablation studies show that Double-Feedback guides LLMs in generating solutions aligned with robotic tasks in the real world.
- 13 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cognitive and Developmental Systems
- Vol. 17 (2) , 287-302
Location Reasoning of Target Objects Based on Human Common Sense and Robot Experiences
The location reasoning of target objects in robotoperated environment is a challenging task. Objects that robots need to interact with are often located at a distance or are contained within containers, making them inaccessible for direct observation by the robot. The uncertainty of the storage location of the target objects and the lacking of the reasoning ability present considerable challenges. In this paper, we propose a method for semantic localization of robot-operated objects based on human common sense and robot experiences. Instead of reasoning the object storage locations solely based on the category of the target object, a probabilistic ontology model is introduced to represent uncertain knowledge in the task of object localization, which combines the expressive power of classical first-order logic and the inference capability of Bayesian inference. The target location is then estimated using the probabilistic ontologies with dynamic integration of human common sense and robot experiences. Experimental results in both simulation and real-world environments demonstrate the effectiveness of the proposed integration of human common sense and robot experiences in the task of semantic localization of robot-operated objects.
- 13 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cognitive and Developmental Systems
- Vol. 17 (2) , 287-302
Location Reasoning of Target Objects Based on Human Common Sense and Robot Experiences
The location reasoning of target objects in robotoperated environment is a challenging task. Objects that robots need to interact with are often located at a distance or are contained within containers, making them inaccessible for direct observation by the robot. The uncertainty of the storage location of the target objects and the lacking of the reasoning ability present considerable challenges. In this paper, we propose a method for semantic localization of robot-operated objects based on human common sense and robot experiences. Instead of reasoning the object storage locations solely based on the category of the target object, a probabilistic ontology model is introduced to represent uncertain knowledge in the task of object localization, which combines the expressive power of classical first-order logic and the inference capability of Bayesian inference. The target location is then estimated using the probabilistic ontologies with dynamic integration of human common sense and robot experiences. Experimental results in both simulation and real-world environments demonstrate the effectiveness of the proposed integration of human common sense and robot experiences in the task of semantic localization of robot-operated objects.
- Open Access
- 29 September 2023
- Published by F1000 Research Ltd in Cobot
- Vol. 1, 13
Research on desktop object grasping based on ellipse fitting
The grasping operation task in service scenes faces several problems including too many kinds of objects and a large amount of training data. This paper focuses on the grasping strategy and pose detection in desktop object grasping tasks. A grasping strategy is given based on the combination of desktop normal vector detection, object category detection, and grasping pose detection. The grasping pose is calculated by ellipse fitting on the depth map. An optimal function is designed to evaluate the possibility of the object sliding along the ellipse axis and the stability of the grasping height. The most reliable grasping pose is selected. Finally, experiments were carried out with a six-degree-of-freedom manipulator, and the proposed grasping method achieved effective grasping of desktop objects without prior knowledge of the object.
- 29 September 2023
- Published by F1000 Research Ltd in Cobot
- Vol. 1, 13
Open AccessResearch on desktop object grasping based on ellipse fitting
The grasping operation task in service scenes faces several problems including too many kinds of objects and a large amount of training data. This paper focuses on the grasping strategy and pose detection in desktop object grasping tasks. A grasping strategy is given based on the combination of desktop normal vector detection, object category detection, and grasping pose detection. The grasping pose is calculated by ellipse fitting on the depth map. An optimal function is designed to evaluate the possibility of the object sliding along the ellipse axis and the stability of the grasping height. The most reliable grasping pose is selected. Finally, experiments were carried out with a six-degree-of-freedom manipulator, and the proposed grasping method achieved effective grasping of desktop objects without prior knowledge of the object.
- Open Access
- 20 May 2022
- Published by F1000 Research Ltd in Cobot
- Vol. 1, 13
Research on desktop object grasping based on ellipse fitting
The grasping operation task in service scenes faces several problems including too many kinds of objects and a large amount of training data. This paper focuses on the grasping strategy and pose detection in desktop object grasping tasks. A grasping strategy is given based on the combination of desktop normal vector detection, object category detection, and grasping pose detection. The grasping pose is calculated by ellipse fitting on the depth map. An optimal function is designed to evaluate the possibility of the object sliding along the ellipse axis and the stability of the grasping height. The most reliable grasping pose is selected. Finally, experiments were carried out with a six-degree-of-freedom manipulator, and the proposed grasping method achieved effective grasping of desktop objects without prior knowledge of the object.
- 20 May 2022
- Published by F1000 Research Ltd in Cobot
- Vol. 1, 13
Open AccessResearch on desktop object grasping based on ellipse fitting
The grasping operation task in service scenes faces several problems including too many kinds of objects and a large amount of training data. This paper focuses on the grasping strategy and pose detection in desktop object grasping tasks. A grasping strategy is given based on the combination of desktop normal vector detection, object category detection, and grasping pose detection. The grasping pose is calculated by ellipse fitting on the depth map. An optimal function is designed to evaluate the possibility of the object sliding along the ellipse axis and the stability of the grasping height. The most reliable grasping pose is selected. Finally, experiments were carried out with a six-degree-of-freedom manipulator, and the proposed grasping method achieved effective grasping of desktop objects without prior knowledge of the object.
- Open Access
- 16 February 2022
- Published by F1000 Research Ltd in Cobot
- Vol. 1
Reasoning about location of robot-operated object based on probabilistic ontologies
Background: A robot needs to acquire the location of objects when performing daily tasks. Compared to industrial robots, a service robot faces a more complex and unstructured working environment where the location of the target object is usually uncertain. For example, due to diversity in personal habits and seasons, apples may be located in refrigerators, tables, or other places in the living environment. Methods: We propose a novel method for semantic localization of the robot-operated object based on probabilistic ontologies (PR-OWL) and multi-entity Bayesian networks (MEBN). The probabilistic web ontology language is used to describe and model the highly uncertain knowledge about the storage location of objects in the human household environment. Furthermore, the target location is inferred based on the multi-entity Bayesian network. Results: The proposed method is capable to adapting to environmental changes and achieves reliable probability estimation of object location. Experiments on simulated robotic tasks verify the effectiveness of the method. Conclusions: We show that applying PR-OWL combined with MEBN to locate the target object for the robot is feasible, which can improve the cognitive and self-adaptive ability of the robot.
- 16 February 2022
- Published by F1000 Research Ltd in Cobot
- Vol. 1
Open AccessReasoning about location of robot-operated object based on probabilistic ontologies
Background: A robot needs to acquire the location of objects when performing daily tasks. Compared to industrial robots, a service robot faces a more complex and unstructured working environment where the location of the target object is usually uncertain. For example, due to diversity in personal habits and seasons, apples may be located in refrigerators, tables, or other places in the living environment. Methods: We propose a novel method for semantic localization of the robot-operated object based on probabilistic ontologies (PR-OWL) and multi-entity Bayesian networks (MEBN). The probabilistic web ontology language is used to describe and model the highly uncertain knowledge about the storage location of objects in the human household environment. Furthermore, the target location is inferred based on the multi-entity Bayesian network. Results: The proposed method is capable to adapting to environmental changes and achieves reliable probability estimation of object location. Experiments on simulated robotic tasks verify the effectiveness of the method. Conclusions: We show that applying PR-OWL combined with MEBN to locate the target object for the robot is feasible, which can improve the cognitive and self-adaptive ability of the robot.
- 6 November 2021
- Published by Springer Nature in International Journal of Automation and Computing
- Vol. 18 (6) , 926-934
Mechanical Design and Dynamic Compliance Control of Lightweight Manipulator
- 6 November 2021
- Published by Springer Nature in International Journal of Automation and Computing
- Vol. 18 (6) , 926-934
Mechanical Design and Dynamic Compliance Control of Lightweight Manipulator