Thomas Tang

Duke University · Bachelor of Science in Computer Science August 2024 – May 2027 (Expected) GPA: 4.0 (Dean’s List) Email: zijia.tang@duke.edu | Phone: +1 (858) 340-5757 Research Areas: Embodied Intelligence, Efficient VLA, Robot Control, AI4Science

Seeking a PhD opportunity in Embodied Intelligence/Robot Learning for the fall of 2027.

Download PDF | Download DOCX

Representative Research and Papers

Robotics/Embodied Intelligence

FLARE: A Failure-Aware Framework for Autonomous Correction and Recovery in Vision-Language-Action Models G. Zhao, Z. Tang, X. Chen, Z. Kuang, Y. Tian, G. Li | CVPR 2026

  • Proposed a Retry-and-Reset recovery framework that enables VLA strategies to self-correct after perturbations, including external disturbances such as strategy failures and human intervention.
  • Addresses the “fragility” of VLA models—tiny perturbations can lead to catastrophic failures—enhancing the robustness of VLA models.
  • Achieved a 84.0% success rate in 9 contact-intensive tasks using RoboMimic, surpassing the previous best method’s 26.2%.

RecoverBench: A Benchmark for Error Recovery in Robotic Manipulation Z. Tang, G. Zhao, X. Chen, et al., G. Li | Under review, 2026

  • Constructed a reproducible robot error recovery benchmark with 12 error skills, 136 error subtypes, and 1,360 verified MuJoCo post-error scenarios across 6 manipulation tasks.
  • Designed an automated error injection mechanism to systematically introduce failures during manipulation, enabling standardized and comprehensive robot error recovery evaluation.
  • Collected and enhanced 1,641 human recovery demonstrations into 10,598 recovery trajectories, supporting downstream recovery strategy learning and evaluation.

Latent Bridge: Accelerating Dual-System VLA Models via Latent Feature Prediction Y. Liu, Y. Li, Z. Tang, et al., Y. Chen, H. Li | arXiv, 2026

  • Developed a latent bridging method that reduces the call to the VLM backbone network by efficiently predicting visual features while maintaining task success rates.
  • Achieved up to 75% reduction in VLM calls, resulting in 1.73 times acceleration on LIBERO and RoboCasa.
  • Verified versatility across different VLA strategies (including GR00T and π series models), demonstrating a generalized method for accelerating VLM robot control.

Other Papers

  • scDrugMap, Nature Communications (Impact Factor: 15.7): Benchmark evaluation of a foundational model for drug response prediction on over 340,000 cells.
  • scPerb, Journal of Advanced Research (Impact Factor: 11.79): Constructed a style transfer VAE for cell perturbation prediction with an R² correlation of 99.5%.
  • PINet, ACM-BCB 2024 Quick Start Report: Constructed a CNN-Transformer model for Alzheimer’s disease MRI prediction with an accuracy rate of 96%.

Research Experience

Research Assistant, Duke Center for Computational Evolutionary Intelligence — Duke University | March 2026 – Present | Supervisors: Professor Yiran Chen, Professor Hai Li

  • Studied the efficient reasoning of dual-system VLA strategies, focusing on latent feature prediction, KV-cache modeling, and backbone network acceleration.
  • Built evaluation pipelines for GR00T and π models on LIBERO and RoboCasa, measuring the trade-off between speed and success rate under different VLM call frequencies.
  • Collaborated with senior researchers in cross-model architecture, system optimization, and embodied intelligence evaluation.

Research Assistant, HCP Laboratory — Sun Yat University | May 2025 – May 2026 | Supervisor: Professor Li Guanbin

  • Studied robot control with failure awareness, focusing on post-error scenario generation, recovery skill classification systems, and reproducible recovery evaluation.
  • Designed and verified robot error scenarios covering grasping failures, object drops, collisions, incorrect object interactions, and trajectory rollback failures.
  • Participated in the generation of recovery data and strategy evaluation pipelines based on VLA manipulation systems.

Paid Research Assistant, Yi Zhang Laboratory — Duke University | August 2024 – May 2025 | Department of Neurosurgery & Department of Biostatistics and Bioinformatics

  • Conducted biomedical AI research in single-cell foundational models, cancer knowledge transfer, and medical image analysis.
  • Participated in the model development, experimental design, and paper writing for AI-for-Science papers.

Research Intern, Song Laboratory — Florida University | May 2023 – August 2024 | Department of Health Outcomes and Biomedical Informatics

  • Developed deep learning models for biomedical prediction tasks, including single-cell perturbation response modeling and disease prediction based on MRI.
  • Published biomedical AI papers as the first author and co-author.

Project Experience

LeHome Challenge: Manipulation of Deformable Objects — Team Lead | ICRA 2026 Workshop · Competition in progress

  • Developed clothing folding strategies in the LeHome HI-FI simulation environment for manipulation of deformable objects under low-cost hardware and Sim-to-Real constraints.

Technical Skills

  • Robotics/Simulation: MuJoCo, robosuite, robomimic, Isaac Sim, MimicGen, LIBERO, RoboCasa
  • VLA/Deep Learning: PyTorch, JAX, LoRA fine-tuning, DiT strategies, Transformers, KV-cache modeling, VLM inference
  • Programming Languages: Python, C++, JAVA, Swift, SQL