cv

General Information

Full Name Xiaopan Zhang (张晓攀)
Date of Birth October 13th 1997
Languages English, Mandarin
Email xzhang877@gatech.edu
Google Scholar Xiaopan Zhang

Education

  • Aug. 2026 - Present

    Atlanta, GA

    Ph.D. Student in Machine Learning
    Georgia Institute of Technology
    • Third-year Ph.D. student, advised by Prof. Jiachen Li. Joined Georgia Tech in Aug. 2026 after transferring from UC Riverside.
    • Research interests include embodied AI agents, multi-robot systems, reinforcement learning in robotics, trustworthy autonomy, and robot navigation.
  • Mar. 2024 - Aug. 2026

    Riverside, CA

    Ph.D. Studies in Electrical Engineering
    University of California, Riverside
  • Jan. 2022 - Dec. 2023

    Los Angeles, CA

    M.S. in Computer Science
    University of Southern California
  • Jan. 2019 - May 2021

    Ann Arbor, MI

    B.S. in Computer Science
    University of Michigan, Ann Arbor

Experience

  • Aug. 2026 - Present
    Graduate Student Researcher
    Georgia Institute of Technology
    • Developing an intervention-guided framework for mobile manipulation that traces failures across navigation, perception, and contact, coordinates physical recovery, and selectively adapts robot skills, residual controllers, and VLA policies to improve long-term autonomy and task performance.
    • Developed object-goal navigation methods using foundation models as reward generators for reinforcement learning, enabling semantic, task-aware, goal-conditioned decisions without continuous high-latency foundation-model queries.
  • Jun. 2026 - Aug. 2026
    Student Associate
    Honda Research Institute USA, Inc. (HRI-US)
    • Developed SituLaMP, an LLM/VLM-based framework for long-horizon multi-robot planning, execution-time failure detection, dynamic replanning, and task reallocation while preserving unaffected plan structure.
    • Implemented PDDL planning, multi-robot simulation, and recovery logic for action, perception, and environmental failures, enabling robust online execution and localized recovery across heterogeneous robot teams.
    • Designed benchmark experiments across multi-robot and embodied-agent tasks to evaluate task completion, robustness, recovery efficiency, generalization, and recovery cost across diverse execution failures.
  • Jul. 2025 - Dec. 2025
    Robotics Research Intern
    AlphaZ, Inc.
    • Developed StratMamba, a dual-stream Mamba-based PPO architecture that separates reactive obstacle processing from strategic goal-directed navigation for efficient LiDAR-based obstacle avoidance in cluttered environments.
    • Built scalable training and evaluation pipelines in IsaacLab and Gazebo for obstacle-rich navigation with quadruped and wheeled robots, and deployed the learned policy on a real Unitree GO1 robot using ROS2, RSLidar-16, VICON localization, and real-time velocity control.
    • Benchmarked against LSTM, Transformer, and vanilla Mamba baselines, improving timeout rate, path efficiency, navigation speed, and sim-to-real robustness. The work resulted in an IROS 2026 paper.
  • Mar. 2024 - Aug. 2026
    Graduate Student Researcher
    Trustworthy Autonomous Systems Laboratory, University of California, Riverside
    • Conducted research on trustworthy embodied and multi-agent robotic systems spanning coordination, task allocation, object-goal navigation, uncertainty calibration, and long-horizon planning, resulting in multiple peer-reviewed publications.
    • Developed object-goal navigation methods using foundation models as reward generators for reinforcement learning, enabling semantic, task-aware, goal-conditioned decisions without continuous high-latency foundation-model queries.
    • Designed heterogeneous multi-agent planning and communication frameworks for long-horizon, multi-objective tasks in partially observable environments, with capability-aware task allocation and calibrated inter-agent information exchange.
    • Proposed a two-stage fine-tuned reasoning model (supervised fine-tuning + reinforcement learning) that directly generates executable PDDL plans, enabling adaptive multi-robot task allocation without heuristic search.
    • Integrated learning-based policies with classical planning and safety constraints to improve robustness, generalization, reliability, and deployment readiness across real-world robotic systems and environments.
  • Jan. 2022 - Dec. 2023
    Graduate Research Assistant
    Dynamic Robotics and Control Laboratory, University of Southern California
    • Contributed to a large-scale multi-robot project that controls multiple quadruped (four-legged) robots for enhanced exploration and task execution.
    • Developed task allocation, path planning, and coordination strategies for coordinated exploration with quadruped robot teams.
    • Conducted extensive testing, validation, and data collection of multi-robot collaboration in simulation and in real-world agricultural environments, including strawberry and raspberry fields.

Research Projects

  • CoReNav: Conformal Calibrated Foundation Model Rewards for Object Navigation
    • Proposed a compute-efficient object-goal navigation framework that combines imitation learning with reinforcement learning fine-tuning driven by foundation-model-based rewards, enabling semantically guided exploration without continuous high-latency model querying.
    • Applied conformal prediction to calibrate foundation-model outputs in the navigation loop, including VLM-based room classification and LLM-generated object relevance priors, producing trust-aware rewards that mitigate hallucinations and stabilize RL fine-tuning.
    • Evaluated on the HM3D Semantics dataset, showing improvements over baselines in success rate, path optimality, and time efficiency in photo-realistic environments.
  • CommCP: Efficient Multi-Agent Coordination via LLM-Based Communication with Conformal Prediction (ICRA 2026)
    • Formulated the information-gathering process for natural-language assignments as a multi-agent multi-task embodied question answering (MM-EQA) problem, where robots collaborate in a shared environment and communicate to exchange information and answers.
    • Proposed CommCP, an LLM-based decentralized communication framework for MM-EQA that uses conformal prediction to calibrate generated messages, reducing distractions to other agents while improving communication reliability and efficiency.
    • Created an MM-EQA benchmark with photo-realistic scenarios from the HM3D dataset and released it to facilitate future research.
  • LaMMA-P: Generalizable Multi-Agent Long-Horizon Task Allocation and Planning with LM-Driven PDDL Planner (ICRA 2025)
    • Introduced LaMMA-P, a framework that integrates LLM reasoning with PDDL planning heuristics to solve long-horizon task planning for heterogeneous robot teams with a flexible number of agents.
    • Developed a modular system design that connects LLMs, PDDL planners, and simulation environments to support flexible task decomposition and efficient sub-task allocation based on robot skills and capabilities.
    • Created MAT-THOR, a benchmark of multi-agent long-horizon tasks based on AI2-THOR with standardized tasks and metrics, achieving state-of-the-art success rate and efficiency.

Selected Publications

  • H. Wu, X. Zhang, K. Sinaei, R. Abnavi, K. Weerakoon, C. Bradley, S. Fakoorian, J. Li, D. Ebeigbe. StratMamba: Strategic and Reactive Stream Partitioning for Path-Efficient LiDAR-Based Obstacle Avoidance. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026.
  • X. Zhang, Z. Wang, Z. Li, J. Yao, J. Li. CommCP: Efficient Multi-Agent Coordination via LLM-Based Communication with Conformal Prediction. IEEE International Conference on Robotics and Automation (ICRA), 2026.
  • M. Gao, X. Li, Q. Yan, X. Sun, X. Zhang, C.-M. Huang, J. Li. GUIDES: Guidance Using Instructor-Distilled Embeddings for Pre-trained Robot Policy Enhancement. IEEE International Conference on Robotics and Automation (ICRA), 2026.
  • X. Zhang, H. Qin, F. Wang, Y. Dong, J. Li. LaMMA-P: Generalizable Multi-Agent Long-Horizon Task Allocation and Planning with LM-Driven PDDL Planner. IEEE International Conference on Robotics and Automation (ICRA), 2025.
  • J. Yao, X. Zhang, Y. Xia, Z. Wang, A. K. Roy-Chowdhury, J. Li. Towards Generalizable Safety in Crowd Navigation via Conformal Uncertainty Handling. Conference on Robot Learning (CoRL), 2025.
  • T. Chakraborty, U. Ghosh, X. Zhang, F. F. Niloy, Y. Dong, J. Li, A. K. Roy-Chowdhury, C. Song. HEAL: An Empirical Study on Hallucinations in Embodied Agents Driven by Large Language Models. Findings of the Association for Computational Linguistics: EMNLP, 2025.
  • The full list is on the publications page.

Technical Skills

Robot Learning
Reinforcement learning and PPO
Imitation learning
Supervised fine-tuning and reinforcement learning for reasoning models
Mamba-based robot policies
Planning and Foundation Models
PDDL planning and task allocation
LLM/VLM-based planning, verification, and reward generation
Conformal prediction and uncertainty calibration
Robotics Platforms and Tools
ROS2, IsaacLab, and Gazebo
Unitree GO1 and real-time velocity control
RSLidar-16 and VICON localization
Embodied AI Benchmarks
HM3D Semantics
AI2-THOR and MAT-THOR
Multi-agent multi-task embodied question answering (MM-EQA)

Other Interests

  • Hobbies: Photography, Travel