CV

Download CV (PDF)

Education

• Nanjing University (Suzhou Campus), Suzhou, China
  • B.Sc. Intelligent Science and Technology (Expected 2027)
  • GPA: 4.48 / 5.00
  • Selected Coursework: Data Mining (95), Database Systems (96), Operating Systems (95), Data Structures & Algorithms (95), Machine Learning (94), Deep Learning (94), Programming Practicum (94.9)

Research Interests

• Embodied AI and robotics — perception, spatial memory, and decision-making for reliable action in unstructured real-world environments.
• Hands-on experience spanning navigation, semantic mapping, and real-robot deployment.
• Seeking opportunities for overseas graduate studies (Fall 2027 entry).

Publications & Preprints

• LaViRA: Language-Vision-Robot Actions Translation for Zero-Shot Vision Language Navigation in Continuous Environments
  • H. Ding, Z. Xu, Y. Fang (Yudong), Y. Wu, Z. Chen, J. Shi, J. Huo, Y. Zhang, Y. Gao

  • ICRA 2026. arXiv:2510.19655

    Project Page

  • Contribution: Led real-robot deployment (on Unitree Go1) and system integration.
• INHerit-SG: Incremental Hierarchical Semantic Scene Graphs with RAG-Style Retrieval
  • Y.T.S. Fang, Z. Shi, J. Qiu, Z. Chen, J. Shi, H. Xu, J. Huo, Y. Gao

  • ICRA 2026 Workshop on Robots Meet Prior Maps, Oral Presentation, Best Presentation Finalist. arXiv:2602.12971

    Project Page

  • Contribution: Independent first author; led whole end-to-end pipeline, paper writing, figures/tables, experiment design, visualization, supplement, hardware/communication bring-up, and real-robot data collection.

Research Experience

Incoming Research Intern (Summer 2026) National University of Singapore (NUS) — Advisor: Prof. Lin Shao

• Incoming summer research in embodied AI.

Undergraduate Researcher (2025–present) Inference & Learning Research Group (led by Prof. Yang Gao), State Key Laboratory for Novel Software Technology, Nanjing University

• Supervised by tenure-track AP Prof. Jieqi Shi, who provided guidance on research design, paper writing, and project execution; collaborated closely on prototyping, integration, and manuscript preparation.

Zero-Shot Vision-and-Language Navigation (LaViRA) (2025.05 – 2025.09)

Advisors: Prof. Jieqi Shi, Prof. Jing Huo

• Addressed trade-offs in existing zero-shot VLN-CE methods between scene generalization and MLLM reasoning utilization.
• Developed a coarse-to-fine hierarchical framework that decomposes actions into language planning for high-level decisions, vision grounding for perceptual integration, and robot control for precise movements.
• Leveraged varying scales of Multimodal Large Language Models (MLLMs) to enhance reasoning, grounding, and navigation efficiency in unseen environments.
• Implemented modular decomposition to maintain transparency and support real-world deployment without prior training.
• Achieved state-of-the-art performance on the VLN-CE benchmark with Success Rate (SR) of 38.3\% and Success weighted by Path Length (SPL) of 28.3\% using Gemini-2.5-Pro, demonstrating superior generalization with +16.1\% SR and +17.7\% SPL over InstructNav; deployed on Unitree Go1 and Agilex Cobot Magic robots. Accepted at ICRA 2026.

Active Exploration with Semantic Map Prediction (SEA) (2025.09 – 2025.10)

Advisors: Prof. Jieqi Shi, Prof. Jing Huo

• Tackled limitations in learning-based exploration methods, including the lack of long-term environmental understanding and efficiency in global awareness.
• Designed an iterative prediction-exploration framework using semantic map prediction to forecast missing areas based on current observations.
• Incorporated an ASC-based local mapper for predictions and confidence estimation, alongside RL-based hierarchical policies for two-stage navigation.
• Developed a confidence-aware full mapper to accumulate and adjust local maps, guiding exploration via differences between predicted and actual maps.
• Deployed and debugged the full pipeline on Habitat datasets, with real-robot integration on Unitree Go1 and Agilex Cobot Magic platforms.

Incremental Hierarchical Semantic Scene Graphs (INHerit-SG) (2025.10 – 2026.02)

Advisors: Prof. Jieqi Shi, Prof. Hao Xu, Prof. Jing Huo

• Resolved misalignments in existing semantic scene graphs, including offline processing, lack of interpretability, and flat structures unsuitable for embodied tasks.
• Constructed a semantic graph evaluation dataset, HM3DSem-SQR, focusing on complex natural language command queries and a human study dataset focusing on semantic accuracy.
• Introduced an online system with a Floor-Room-Area-Object hierarchy and RAG-style retrieval, using natural-language descriptions as semantic anchors for human-intent alignment.
• Employed an asynchronous dual-process architecture to decouple geometric segmentation from semantic reasoning, with event-triggered updates for long-term consistency and low overhead.
• Deployed multi-role Large Language Models (LLMs) to decompose queries into atomic constraints, handle logical negations, and apply hard-to-soft filtering for robust reasoning.
• Achieved state-of-the-art results on the HM3DSem-SQR dataset with geometric accuracy of 37.7\% (within 1m, +15\% over DualMap) and semantic accuracy of 70.6\%, real-world trajectory evaluations with 60.0\% success rate (+70\% over baselines). The system also demonstrates scalability for downstream navigation tasks.

Research in Progress

Lifelong Mobile-Manipulation Memory Benchmark (2026.03 – present)

• Designing a benchmark that evaluates world-model maintenance under continuous household perturbations, organized around a three-level capability framework (build/organize world model, update under conflicts, memory-guided action). Simulation environment implemented; preliminary method under active development.

Uncertainty-Aware Interactive Object-Goal Navigation (2026.06 – present)

• Developing a decision framework for when an embodied agent should act, explore, or ask the user, based on two-layer uncertainty modeling and information-gain-driven question generation. Research proposal completed; feasibility validated via baseline (SG-Nav, CoIN-Bench) reproduction.

Honors & Awards

• First Prize, NJU Scholarship for HK/Macao & Overseas Chinese Students (Sophomore year; university-wide; 5 awardees (Top 2%); defense required)
• Third Prize, NJU Scholarship for HK/Macao & Overseas Chinese Students (Freshman year; university-wide)

Academic Service

• Reviewer (by invitation), IEEE Robotics and Automation Letters (RA-L), 2026

Skills

• Programming: Python (for ML pipelines and ROS integration), C++
• Robotics: ROS (for data collection across diverse sensors, inter-device communication in complex setups, downstream control via SDKs for dynamic motion planning and execution), real-robot deployment and debugging
• Tools: Linux, Git; agentic coding workflows (Claude Code / Codex) in remote development for rapid prototyping and project automation
• Hardware: 3D modeling & printing (custom fixtures for sensor integration); sensor/compute configurations (possess hands-on experience with various commonly used robot cameras)