About Me

I am a master's student in Computational Science and Engineering at Harvard University, working with Prof. Mengyu Wang (Harvard AI and Robotics Lab) and Prof. Paul Pu Liang (MIT Media Lab). My research sits at the intersection of robotics, embodied AI, and 3D vision.

I received my B.S.E. in Computer Engineering from the University of Michigan, where I worked with Prof. Ram Vasudevan in the ROAHM Lab on differentiable models of deformable linear objects and multi-modal Bayesian perception.

My work builds physically grounded models that robots can actually act on — from real-time differentiable simulators for deformable objects, to perception and action that fuse vision and touch, to geometry-aware video world models for manipulation. I'm always open to research discussions and collaboration. Feel free to reach out via yuzhenchen@g.harvard.edu.

Research Interests

My goal is to give robots a physically grounded understanding of the world, the ability to reason about how it will change, and the means to act on it alongside people. My research spans three connected threads:

  • Predictive World Models: Building models that can imagine plausible futures — anticipating how a scene will change and what people intend — so robots can reason and act reliably over long horizons.
  • Human-Interactive, Agentic Robotics: Moving robots beyond one-shot instruction following toward agents that interact with people as fluently as language models do in software — able to explain their decisions, and to be redirected, corrected, and taught through natural interaction.
  • Active Perception & Embodied Understanding: Robots that do not just perceive passively, but actively decide where to look and how to move to understand a dynamic 3D world — its geometry, motion, and interactions — and act robustly within it.

News

  • [Oct. 2026] âś‹ T-WLA is on arXiv — a tactile world–language–action model bridging slow vision and fast touch.
  • [Oct. 2026] 🎉 DEFT is accepted by IJRR (The International Journal of Robotics Research).
  • [Sep. 2026] 🎥 WCM demo is accepted by the NeurIPS 2026 Embodied Spatial Reasoning (ESR) Workshop Demo Track.
  • [Aug. 2026] 🤝 Serving as a Junior Organizer of the 1st Workshop on Physical World AI (PhysWorldAI) at NeurIPS 2026.
  • [Jul. 2026] đź“„ Learning Ophthalmologist Clinical Reasoning for Glaucoma Diagnosis is out on medRxiv.
  • [Jul. 2026] 🎉 Act2See is accepted by the RSS 2026 FM4RoboPlan Workshop.
  • [Jul. 2026] 🎉 World Cognition Model (WCM) is accepted by the RSS 2026 Workshop on Human-centric Mobile Manipulation.
  • [Jun. 2026] 🎓 Awarded the OMPP Summer Research Fellowship at Harvard University.
  • [Jun. 2026] 🎉 GEM-4D is accepted by ECCV 2026.
  • [Jun. 2026] 🏆 GEM-4D — Best Paper Award, Runner-Up at the CVPR 2026 WMAS Workshop.
  • [Jun. 2026] ✨ GEM-4D — Spotlight at the CVPR 2026 3D-LLM/VLA Workshop.
  • [Mar. 2026] 🎉 SLIM-VDB is accepted by IEEE RA-L.
  • [May. 2025] đź“Š MatPredict — a dataset and benchmark for learning material properties of indoor objects — is released.
  • [Feb. 2025] 🤖 DEFT is on arXiv — differentiable branched discrete elastic rods for furcated DLOs.
  • [Sep. 2024] 🤖 DEER is accepted by CoRL 2024.
  • [Apr. 2024] âś‹ You've Got to Feel It to Believe It is accepted by RSS 2024.
  • [Oct. 2023] 🏆 Multi-Modal Semantic Perception Using Bayesian Inference — Best Poster at the IROS 2023 Workshop.

Publications

* denotes equal contribution.

  1. Teaser for T-WLA

    T-WLA: Bridging Slow Vision and Fast Touch

    Kaichen Zhou*, Yuzhen Chen*, Hannah Yao*, Chonghe Jiang, Rao Fu, Xihang Yu, Wei Dai, Bashar Kabbarah, Haoyu Chen, Ishita Goluguri, Zhiyang Dou, Haonan Chen, Fangneng Zhan, Mengyu Wang, Wojciech Matusik, Antonio Torralba, Paul Pu Liang

    A multi-rate world–language–action architecture that bridges low-frequency visual planning and high-frequency tactile correction through shared visual memory: a lightweight tactile editor uses cached task context and recent touch to asynchronously refine unexecuted actions without rerunning vision, with future-image and pressure prediction as training-only supervision. Cuts finger-joint MPJPE by 30% on EgoTouch and 27% on Tachin, and reaches 50% average success on four real-world dexterous manipulation tasks (vs. 23% for T-Rex).

    arXiv preprint · 2026 Robotics
    Paper Project
  2. Teaser for GEM-4D

    GEM-4D: Geometry-Enhanced Video World Models for Robot Manipulation

    Kaichen Zhou*, Yuzhen Chen*, Fangneng Zhan, Hang Hua, Grace Chen, Xinhai Chang, Ao Qu, Yilun Du, Zhuang Liu, Paul Pu Liang, Mengyu Wang

    A geometry-grounded video world model that distills dense 4D correspondence supervision into a video diffusion transformer during training (no added inference cost), paired with an adaptive inverse-dynamics system that turns generated videos into executable 6-DoF robot trajectories. State-of-the-art 4D scene generation (SSIM 76.2→82.1, FVD 33.2→31.8, Chamfer 0.244→0.200) and real-world manipulation success lifted from 61% to 81%. Best Paper Runner-Up (CVPR 2026 WMAS Workshop); Spotlight (CVPR 2026 3D-LLM/VLA Workshop).

    ECCV · 2026 Robotics
    Paper Project Poster
  3. Teaser for WCM

    WCM: World-Cognition Model for Generalizable Human-Robot Interaction

    Yuzhen Chen, Kaichen Zhou

    A human-centered embodied agent built on the SLAK architecture (Sensing, Logic, Action, Knowledge) with an asynchronous runtime, so reasoning, dialogue, and execution proceed concurrently instead of halting to think between steps. A human-in-the-loop teaching mode and a chain-of-thought distillation pipeline let the agent learn long-horizon tasks from user demonstration and keep improving. 73.8% average success across a nine-task real-world human–robot interaction suite, with task time cut from 103.5 s to 38.5 s against a VLA baseline. Demo accepted to the NeurIPS 2026 ESR Workshop Demo Track.

    RSS 2026 Workshop on Human-centric Mobile Manipulation · 2026 Robotics
    Paper Project Poster Demo videos
  4. Teaser for Act to See

    Act to See: Evolving Articulated Memory for Closed-Loop Vision-Language Planning

    Yifeng Liu*, Yuzhen Chen*, Bingyang Wang*, Chengqi Li, Jingyi Lu, Han Li, Y. Li, M. Ji, Kaichen Zhou, Mengyu Wang, Fangneng Zhan

    A VLM-guided manipulation framework that keeps an evolving URDF as spatial memory, alongside task progress and interaction outcomes, to decide when to probe, manipulate, or stop. Each interaction estimates joint parameters from observed part motion, fuses pose-aligned RGB-D geometry, and uses the updated model to generate the next linear or revolute trajectory. 100% valid joint reconstruction in 4.23 s and 81.8% drawer / 72.7% door completion in SAPIEN; 5/5 drawer and 4/5 door successes on a UF850 robot in fixed setups.

    Manuscript · 2026 Robotics
    Project
  5. Teaser for Glaucoma Reasoning

    Learning Ophthalmologist Clinical Reasoning for Glaucoma Diagnosis from Fundus Images

    Kaichen Zhou*, Yuzhen Chen*, Elif Yildiz*, Min Shi, David Dai, Grace Chen, Jiale Zheng, He Wang, Fangneng Zhan, Chhavi Saini, Lucy Q. Shen, Yike Guo, Paul Pu Liang, Mengyu Wang

    A two-stage vision–language reasoning pipeline that couples a frozen RETFound encoder with MedGemma-27B: attention pooling and five task-specific evidence heads, trained with uncertainty-weighted multi-task objectives, extract structured visual evidence that conditions six-step chain-of-thought generation alongside the fundus image. MedGemma-27B is adapted with LoRA and supervised CoT fine-tuning (class-weighted next-token loss on response tokens only), and a five-fold out-of-fold evidence-generation and report-adaptation pipeline reduces training leakage while exposing the reasoning model to realistic prediction errors. 94.7% balanced accuracy, 94.8% precision, and 0.874 BERTScore-F1 against expert reports.

    medRxiv preprint · 2026 · under review at a journal Medical AI
    Paper

Education

Harvard University

Harvard University, Cambridge, MA

Sept. 2025 – May 2027 (Expected)

M.S. in Computational Science and Engineering

Coursework: Video Generation, Multi-modal World Models & VLA.

University of Michigan

University of Michigan, Ann Arbor, MI

Aug. 2021 – May 2025

B.S.E. in Computer Engineering · GPA 3.9 / 4.0

Machine Learning, Computer Vision, SLAM, Robot Perception, GPU Parallel Computation, Robot 3D Reconstruction.

Experience

Harvard AI and Robotics Lab

Harvard AI and Robotics Lab

Sept. 2025 – Present

Research Assistant · advised by Prof. Mengyu Wang

Tactile world–language–action models (T-WLA); geometry-enhanced video world models for robot manipulation (GEM-4D); evolving articulated memory for VLM-guided manipulation (Act to See); world-cognition model for human–robot interaction (WCM, with Dr. Kaichen Zhou); ophthalmologist-style clinical reasoning with retinal foundation models and medical VLMs (glaucoma diagnosis).

MIT Media Lab

Multisensory Intelligence Lab, MIT Media Lab

Sept. 2025 – Present

Research Assistant · advised by Prof. Paul Pu Liang

Tactile world–language–action models (T-WLA, with Prof. Antonio Torralba and Prof. Wojciech Matusik); geometry-enhanced video world models for robot manipulation (GEM-4D); ongoing spatially calibrated visuotactile data collection for dexterous manipulation.

ROAHM Lab

ROAHM Lab

Apr. 2023 – May 2025

Research Assistant · advised by Prof. Ram Vasudevan

Real-time differentiable modeling of deformable linear objects (DEER, DEFT); multi-modal Bayesian semantic perception; conformalized safe motion planning (CROWS); material-property learning from images (MatPredict, with Prof. Arpan Kusari).

Honors & Awards

  • [2026] 🎓 OMPP Summer Research Fellowship, Harvard University.
  • [2026] 🏆 Best Paper Award, Runner-Up — CVPR 2026 WMAS Workshop.
  • [2026] ✨ Spotlight — CVPR 2026 3D-LLM/VLA Workshop.
  • [2023] 🥇 Best Poster Award — IROS 2023 IPPC Workshop (Integrated Perception, Planning, and Control).
  • [2023–24] 🎖️ James B. Angell Scholar, University of Michigan.
  • [2022–25] đź“ś Dean's List, University of Michigan.

Services & Teaching

  • [2026] Junior Organizer, PhysWorldAI: 1st Workshop on Physical World AI, NeurIPS 2026.
  • [2026–] Robotics Community Event Organizer, MIT CHIEF.
  • [2022–25] Teaching Assistant, EECS 442 (Computer Vision) & Physics 140X, University of Michigan.
  • [2023–24] Departmental Ambassador, ENGR 110 & UMich Robotics Department.
  • [2023] Student Tutor, UMich Engineering Center for Academic Success.