Xian Guo
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Xian Guo has a total of 62 co-authors. Here are the top co-authors by number of shared publications:
- Yongchun Fang
Institute of Robotics and Automatic Information Systems, College of Artificial Intelligence, Nankai University, Tianjin, China
11 shared publications - Huawang Liu
Institute of Robotics and Automatic Information System, College of Artificial Intelligence, Nankai University, Tianjin, China
5 shared publications - Wei Huang
Institute of Robotics and Automatic Information System, College of Artificial Intelligence, Nankai University, Tianjin, China
5 shared publications - Lixing Liu
Institute of Robotics and Automatic Information System, College of Artificial Intelligence, Nankai University, Tianjin, China
4 shared publications - Fuxi Wan
Institute of Robotics and Automatic Information System, College of Artificial Intelligence, Nankai University, Tianjin, China
3 shared publications - Wei Zhu
Georgia Institute of Technology
3 shared publications
Latest Publications
- Open Access
- 29 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 140-148
SC2-VLA: A Vision-Language-Action Model for Strategic Decision Making in StarCraft II
- 29 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 140-148
Open AccessSC2-VLA: A Vision-Language-Action Model for Strategic Decision Making in StarCraft II
- 13 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Robotics and Automation Letters
- Vol. 11 (5) , 5646-5653
Whole-Body Path Following of an Active-Joint Active-Wheel Snake Robot Based on Head-Body Stepwise Control
This letter presents a two-dimensional planar path-following controller for active-joint active-wheel snake-like robots. Compared to passive-wheel snake robots, active-wheel snake robots can navigate narrower spaces and generate stronger driving forces, enabling adaptation to various terrains. However, due to the highly redundant multi-link and multi-active-wheel body configuration, achieving whole-body path following for active-joint active-wheel snake robots is challenging. To address this, we propose a head-body stepwise path-following control method based on a multi-rigid-body kinematic model to solve the problem of high degrees of freedom and the coordination between joints and wheels. First, by applying the no-side-slip constraint, we construct a multi-joint and multi-wheel kinematic model for the snake body path-following control task. Then, a motion controller is designed for the fundamental motion of the active wheeled snake robot. To achieve path following, an arbitrary path following controller is designed for turning angle. Then, the designed turning angle is substituted into the motion controller, and the angular velocity of each joint is calculated based on kinematic model. Subsequently, an incremental PID controller is designed to control the angle of each joint. Finally, the whole-body path following task of active-wheel snake-like robot can be implemented. Simulation and extensive experimental results for both curvature-continuous and curvature-discontinuous paths demonstrate the superior following performance of the proposed controller compared to the PD method and passive-joint conditions. Furthermore, the robustness of the proposed method to variations in friction is validated.
- 13 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Robotics and Automation Letters
- Vol. 11 (5) , 5646-5653
Whole-Body Path Following of an Active-Joint Active-Wheel Snake Robot Based on Head-Body Stepwise Control
This letter presents a two-dimensional planar path-following controller for active-joint active-wheel snake-like robots. Compared to passive-wheel snake robots, active-wheel snake robots can navigate narrower spaces and generate stronger driving forces, enabling adaptation to various terrains. However, due to the highly redundant multi-link and multi-active-wheel body configuration, achieving whole-body path following for active-joint active-wheel snake robots is challenging. To address this, we propose a head-body stepwise path-following control method based on a multi-rigid-body kinematic model to solve the problem of high degrees of freedom and the coordination between joints and wheels. First, by applying the no-side-slip constraint, we construct a multi-joint and multi-wheel kinematic model for the snake body path-following control task. Then, a motion controller is designed for the fundamental motion of the active wheeled snake robot. To achieve path following, an arbitrary path following controller is designed for turning angle. Then, the designed turning angle is substituted into the motion controller, and the angular velocity of each joint is calculated based on kinematic model. Subsequently, an incremental PID controller is designed to control the angle of each joint. Finally, the whole-body path following task of active-wheel snake-like robot can be implemented. Simulation and extensive experimental results for both curvature-continuous and curvature-discontinuous paths demonstrate the superior following performance of the proposed controller compared to the PD method and passive-joint conditions. Furthermore, the robustness of the proposed method to variations in friction is validated.
- 16 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Electronics
- Vol. 73 (6) , 8909-8921
A Head-Orientation Stabilized Gait With 3-D Backbone Curve Shape and Raised Head for Snake Robots
Head stabilization is crucial for 3-D snake robots to perceive their environment effectively using onboard vision sensors. This article proposes a head-orientation-stabilized gait that enables the snake robot to capture wide-field video with improved stability. Given the challenge of compensating for coupled swings in both directions, the snake robot is divided into a motion part and a head-raising part, as guided by the proposed design rules. The motion section suppresses body-induced disturbances while preserving off-road locomotion capability and the head-raising section uses a hook shape to broaden the camera view and reduce residual disturbances. To generate the motion part, the S-pedal gait, suited for uneven terrain, is reformulated into Frenet–Serret-based angle functions for localized body tuning without full-shape transmission. Furthermore, setting specific rolling angles leverages segment twist angles to decouple 3-D shape formation, where pitch and yaw joints generate floating and grounding arcs, respectively, enabling independent joint modulation and enhanced swing suppression. A series of experiments validates the proposed gait, with results showing satisfactory head stabilization, wider camera coverage, and improved adaptability to challenging terrain compared to existing methods.
- 16 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Electronics
- Vol. 73 (6) , 8909-8921
A Head-Orientation Stabilized Gait With 3-D Backbone Curve Shape and Raised Head for Snake Robots
Head stabilization is crucial for 3-D snake robots to perceive their environment effectively using onboard vision sensors. This article proposes a head-orientation-stabilized gait that enables the snake robot to capture wide-field video with improved stability. Given the challenge of compensating for coupled swings in both directions, the snake robot is divided into a motion part and a head-raising part, as guided by the proposed design rules. The motion section suppresses body-induced disturbances while preserving off-road locomotion capability and the head-raising section uses a hook shape to broaden the camera view and reduce residual disturbances. To generate the motion part, the S-pedal gait, suited for uneven terrain, is reformulated into Frenet–Serret-based angle functions for localized body tuning without full-shape transmission. Furthermore, setting specific rolling angles leverages segment twist angles to decouple 3-D shape formation, where pitch and yaw joints generate floating and grounding arcs, respectively, enabling independent joint modulation and enhanced swing suppression. A series of experiments validates the proposed gait, with results showing satisfactory head stabilization, wider camera coverage, and improved adaptability to challenging terrain compared to existing methods.
- Open Access
LLM-Cave: A benchmark and light environment for large language models reasoning and decision-making system
Large language models (LLMs) such as ChatGPT o1, ChatGPT o3, and DeepSeek R1 have shown great potential in solving difficult problems. However, current LLM evaluation benchmarks are limited to one-step interactions. Some of the existing sequence decision-making environments, such as TextStarCraftII and LLM-PySC2, are too complicated and require hours of interaction to complete a game. In this paper, we introduce LLM-Cave, a benchmark and light environment for LLM reasoning and decision-making systems. This environment is a classic instance in the era of Symbolism. Artificial intelligence enables the agent to explore the environment and avoid potential losses by reasoning about nearby dangers using partial observable state information. In the experiment, we evaluated the sequential reasoning ability, decision-making performance and computational efficiency of mainstream large language models (LLMs) such as GPT-4o-mini, o1-mini, and DeepSeek-R1. Experiments show that while Deepseek-R1 achieved the highest success rate on complex reasoning tasks, smaller models like 4o-mini significantly narrowed the performance gap on challenges by employing Chain of Speculation and Planner-Critic strategies, at the expense of reduced computational efficiency. This indicates that structured, multi-step reasoning combined with an LLM-based feedback mechanism can substantially enhance an LLM's decision-making capabilities, providing a promising direction for improving reasoning in weaker models and suggesting a new reasoning-centered benchmark for LLM assessment. Our code is open-sourced in https://github.com/puleya1277/CaveEnv.
Open AccessLLM-Cave: A benchmark and light environment for large language models reasoning and decision-making system
Large language models (LLMs) such as ChatGPT o1, ChatGPT o3, and DeepSeek R1 have shown great potential in solving difficult problems. However, current LLM evaluation benchmarks are limited to one-step interactions. Some of the existing sequence decision-making environments, such as TextStarCraftII and LLM-PySC2, are too complicated and require hours of interaction to complete a game. In this paper, we introduce LLM-Cave, a benchmark and light environment for LLM reasoning and decision-making systems. This environment is a classic instance in the era of Symbolism. Artificial intelligence enables the agent to explore the environment and avoid potential losses by reasoning about nearby dangers using partial observable state information. In the experiment, we evaluated the sequential reasoning ability, decision-making performance and computational efficiency of mainstream large language models (LLMs) such as GPT-4o-mini, o1-mini, and DeepSeek-R1. Experiments show that while Deepseek-R1 achieved the highest success rate on complex reasoning tasks, smaller models like 4o-mini significantly narrowed the performance gap on challenges by employing Chain of Speculation and Planner-Critic strategies, at the expense of reduced computational efficiency. This indicates that structured, multi-step reasoning combined with an LLM-based feedback mechanism can substantially enhance an LLM's decision-making capabilities, providing a promising direction for improving reasoning in weaker models and suggesting a new reasoning-centered benchmark for LLM assessment. Our code is open-sourced in https://github.com/puleya1277/CaveEnv.
- Open Access
FRMA: Four-Phase Rapid Motor Adaptation Framework
In many real-world control tasks, agents operate under partial observability, where access to complete state information is limited or corrupted by noise. This poses significant challenges for reinforcement learning algorithms, as methods relying on full states or long observation histories can be computationally expensive and less robust. Four-Phase Rapid Motor Adaptation (FRMA) is a reinforcement learning framework designed to address these challenges in high-frequency control tasks under partial observability. FRMA proceeds through four sequential stages: (i) full-state pretraining to establish a strong initial policy, (ii) auxiliary hidden-state prediction for LSTM memory initialization, (iii) aligned latent representation learning to bridge partial observations with full-state dynamics, and (iv) latent-state policy fine-tuning for robust deployment. Notably, FRMA leverages full-state information () only during training to supervise latent representation learning, while at deployment it requires only short sequences of recent observations and actions. This allows agents to infer compact and informative latent states, achieving performance comparable to policies with full-state access. Extensive experiments on continuous control benchmarks show that FRMA attains near-optimal performance even with minimal observation–action histories, reducing reliance on long-term memory and computational resources. Moreover, FRMA demonstrates strong robustness to observation noise, maintaining high control accuracy under substantial sensory corruption. These results indicate that FRMA provides an effective and generalizable solution for partially observable control tasks, enabling efficient and reliable agent operation when full state information is unavailable or noisy.Open AccessFRMA: Four-Phase Rapid Motor Adaptation Framework
In many real-world control tasks, agents operate under partial observability, where access to complete state information is limited or corrupted by noise. This poses significant challenges for reinforcement learning algorithms, as methods relying on full states or long observation histories can be computationally expensive and less robust. Four-Phase Rapid Motor Adaptation (FRMA) is a reinforcement learning framework designed to address these challenges in high-frequency control tasks under partial observability. FRMA proceeds through four sequential stages: (i) full-state pretraining to establish a strong initial policy, (ii) auxiliary hidden-state prediction for LSTM memory initialization, (iii) aligned latent representation learning to bridge partial observations with full-state dynamics, and (iv) latent-state policy fine-tuning for robust deployment. Notably, FRMA leverages full-state information () only during training to supervise latent representation learning, while at deployment it requires only short sequences of recent observations and actions. This allows agents to infer compact and informative latent states, achieving performance comparable to policies with full-state access. Extensive experiments on continuous control benchmarks show that FRMA attains near-optimal performance even with minimal observation–action histories, reducing reliance on long-term memory and computational resources. Moreover, FRMA demonstrates strong robustness to observation noise, maintaining high control accuracy under substantial sensory corruption. These results indicate that FRMA provides an effective and generalizable solution for partially observable control tasks, enabling efficient and reliable agent operation when full state information is unavailable or noisy. - 19 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ASME Transactions on Mechatronics
- Vol. 31 (2) , 1558-1569
Autonomous Visual Navigation With Head Stabilization Control for a Salamander-Like Robot
Salamander-like robots, designed to replicate the biological features and movement principles of real salamanders, offer impressive agility and adaptability for autonomous tasks, such as environmental monitoring and disaster relief. However, their application in visual navigation within real-world scenarios remains largely unexplored. In this article, we propose a novel visual navigation framework for salamander-like robots, combining head stabilization, obstacle detection, and fuzzy control into a cohesive system. The head stabilization controller ensures stable visual perception by reducing the periodic yaw motion using a Fourier series, with neural networks estimating the coefficients and maintaining roll and pitch stability by calculating leg joint adjustments based on attitude feedback. In addition, an innovative attention-based neural network is introduced to improve obstacle detection accuracy by focusing on critical visual features, and a hierarchical fuzzy control scheme is established to generate locomotion control signals for obstacle avoidance and target navigation. Extensive experiments demonstrate the effectiveness of the framework in visual stabilization, obstacle detection, and target navigation, highlighting its practical applicability for a salamander-like robot in real-world environments.
- 19 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ASME Transactions on Mechatronics
- Vol. 31 (2) , 1558-1569
Autonomous Visual Navigation With Head Stabilization Control for a Salamander-Like Robot
Salamander-like robots, designed to replicate the biological features and movement principles of real salamanders, offer impressive agility and adaptability for autonomous tasks, such as environmental monitoring and disaster relief. However, their application in visual navigation within real-world scenarios remains largely unexplored. In this article, we propose a novel visual navigation framework for salamander-like robots, combining head stabilization, obstacle detection, and fuzzy control into a cohesive system. The head stabilization controller ensures stable visual perception by reducing the periodic yaw motion using a Fourier series, with neural networks estimating the coefficients and maintaining roll and pitch stability by calculating leg joint adjustments based on attitude feedback. In addition, an innovative attention-based neural network is introduced to improve obstacle detection accuracy by focusing on critical visual features, and a hierarchical fuzzy control scheme is established to generate locomotion control signals for obstacle avoidance and target navigation. Extensive experiments demonstrate the effectiveness of the framework in visual stabilization, obstacle detection, and target navigation, highlighting its practical applicability for a salamander-like robot in real-world environments.
- Open Access
Hierarchical Expert Prompt for Large-Language-Model: An Approach Defeat Elite AI in TextStarCraft II for the First Time
Since the emergence of the Large Language Model (LLM), LLM has been widely used in fields such as writing, translating, and searching. However, there is still great potential for LLM-based methods in handling complex tasks such as decision-making in the StarCraft II environment. To address problems such as lack of relevant knowledge and poor control over subtasks of varying importance, we propose a Hierarchical Expert Prompt (HEP) for LLM. Our method improves the understanding of game situations through expert-level tactical knowledge, improving the processing quality of tasks of varying importance through a hierarchical framework. Our approach defeated the highest level (Elite) standard built-in agent in TextStarCraft II for the first time and consistently outperformed the baseline method in other difficulties. Our experiments suggest that the proposed method is a practical solution for tackling complex decision-making challenges. The replay video can be viewed on https://www.bilibili.com/video/BV1uz42187EF and https://youtu.be/dO3PshWLV5M, and our codes have been open-sourced on https://github.com/luchang1113/HEP-LLM-play-StarCraftII.
Open AccessHierarchical Expert Prompt for Large-Language-Model: An Approach Defeat Elite AI in TextStarCraft II for the First Time
Since the emergence of the Large Language Model (LLM), LLM has been widely used in fields such as writing, translating, and searching. However, there is still great potential for LLM-based methods in handling complex tasks such as decision-making in the StarCraft II environment. To address problems such as lack of relevant knowledge and poor control over subtasks of varying importance, we propose a Hierarchical Expert Prompt (HEP) for LLM. Our method improves the understanding of game situations through expert-level tactical knowledge, improving the processing quality of tasks of varying importance through a hierarchical framework. Our approach defeated the highest level (Elite) standard built-in agent in TextStarCraft II for the first time and consistently outperformed the baseline method in other difficulties. Our experiments suggest that the proposed method is a practical solution for tackling complex decision-making challenges. The replay video can be viewed on https://www.bilibili.com/video/BV1uz42187EF and https://youtu.be/dO3PshWLV5M, and our codes have been open-sourced on https://github.com/luchang1113/HEP-LLM-play-StarCraftII.
- 19 December 2024
- Published by Elsevier in Expert Systems with Applications
- Vol. 267
Multi-UAV air combat cooperative game based on virtual opponent and value attention decomposition policy gradient
- 19 December 2024
- Published by Elsevier in Expert Systems with Applications
- Vol. 267
Multi-UAV air combat cooperative game based on virtual opponent and value attention decomposition policy gradient
- 25 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Robotics and Automation Letters
- Vol. 10 (1) , 311-318
A Maneuverable Winding Gait for Snake Robots Based on a Delay-Aware Swing and Grasp Framework Combining Rules and Learning Methods
Due to the high redundant degree of freedom characteristics of snake robots, their joint lever arms tend to be very long and often result in torque saturation, especially in the case of inter-tree motion. Traditional static planning methods based on curve segments connecting or wave functions are often limited by torque saturation and cannot meet the requirements of inter-tree motion of snake robots. Therefore, in this letter, a delay-aware swing and grasp framework combining rules and learning methods (DSG) is proposed for extending the inter-tree motion capability of snake robots. Specifically, first, to overcome the torque saturation problem, a joint torque direction determination rule is proposed to fully utilize the kinetic energy of the snake robot and enable the robot to move in the desired manner. Then, the DSG reduces the exploration space of the policy and guarantees the performance under delay, and based on which a maneuverable winding gait is designed to enable it to wrap around the target horizontal branch with maneuverability. Simulation and sufficient experiment results demonstrate that the proposed reinforcement learning (RL) controller has low torque requirement, high robustness under high delay, fast motion speed, and high generalizability.
- 25 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Robotics and Automation Letters
- Vol. 10 (1) , 311-318
A Maneuverable Winding Gait for Snake Robots Based on a Delay-Aware Swing and Grasp Framework Combining Rules and Learning Methods
Due to the high redundant degree of freedom characteristics of snake robots, their joint lever arms tend to be very long and often result in torque saturation, especially in the case of inter-tree motion. Traditional static planning methods based on curve segments connecting or wave functions are often limited by torque saturation and cannot meet the requirements of inter-tree motion of snake robots. Therefore, in this letter, a delay-aware swing and grasp framework combining rules and learning methods (DSG) is proposed for extending the inter-tree motion capability of snake robots. Specifically, first, to overcome the torque saturation problem, a joint torque direction determination rule is proposed to fully utilize the kinetic energy of the snake robot and enable the robot to move in the desired manner. Then, the DSG reduces the exploration space of the policy and guarantees the performance under delay, and based on which a maneuverable winding gait is designed to enable it to wrap around the target horizontal branch with maneuverability. Simulation and sufficient experiment results demonstrate that the proposed reinforcement learning (RL) controller has low torque requirement, high robustness under high delay, fast motion speed, and high generalizability.
- 16 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ASME Transactions on Mechatronics
- Vol. 30 (2) , 851-861
Representation Reinforcement Learning-Based Dense Control for Point Following With State Sparse Sensing of 3-D Snake Robots
During robot movements, the environmental states often fail to update in real-time due to interference from various factors, such as obstacle obstructions, communication disruptions, etc., which commonly results in interruptions or even failures in motion control. To achieve dense motion control under sparse state sensing, an important challenge is to predict future multiple actions based on sparse states, which is hindered by the large and complex action search space. Unfortunately, limited research has been dedicated to addressing this challenge. Therefore, this article proposes a representation reinforcement learning (RRL) based solution, called Sparse-State to Dense-Actions Latent Control , designed to realize dense motion control of 3-D snake robots subject to sparse environmental state sensing, which guarantees satisfactory point following performance. In particular, by introducing a latent representation of multiple actions, the control policy optimizes latent actions to predict dense motion gaits, which significantly enhances training efficiency and motion performance. Meanwhile, to learn a compact latent variable model, three mechanisms are proposed to ensure its efficient training, semantic smoothness, and energy efficiency, facilitating exploration of the RL algorithm. To the best of our knowledge, this article provides the first solution that enables a 3-D snake robot to successfully accomplish point following tasks under sparse state sensing. Simulation and practical experiments confirm the effectiveness, robustness, and generalizability of the proposed algorithm, with all following errors less than 0.02 m.
- 16 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ASME Transactions on Mechatronics
- Vol. 30 (2) , 851-861
Representation Reinforcement Learning-Based Dense Control for Point Following With State Sparse Sensing of 3-D Snake Robots
During robot movements, the environmental states often fail to update in real-time due to interference from various factors, such as obstacle obstructions, communication disruptions, etc., which commonly results in interruptions or even failures in motion control. To achieve dense motion control under sparse state sensing, an important challenge is to predict future multiple actions based on sparse states, which is hindered by the large and complex action search space. Unfortunately, limited research has been dedicated to addressing this challenge. Therefore, this article proposes a representation reinforcement learning (RRL) based solution, called Sparse-State to Dense-Actions Latent Control , designed to realize dense motion control of 3-D snake robots subject to sparse environmental state sensing, which guarantees satisfactory point following performance. In particular, by introducing a latent representation of multiple actions, the control policy optimizes latent actions to predict dense motion gaits, which significantly enhances training efficiency and motion performance. Meanwhile, to learn a compact latent variable model, three mechanisms are proposed to ensure its efficient training, semantic smoothness, and energy efficiency, facilitating exploration of the RL algorithm. To the best of our knowledge, this article provides the first solution that enables a 3-D snake robot to successfully accomplish point following tasks under sparse state sensing. Simulation and practical experiments confirm the effectiveness, robustness, and generalizability of the proposed algorithm, with all following errors less than 0.02 m.