Yi (Jasper) Zhao

Master's Student · AirLab · Robotics Institute, Carnegie Mellon University

I am a Master's Student at AirLab, Carnegie Mellon University Robotics Institute, pursuing an M.S. in Computer Vision (MSCV) (Aug 2026 – Dec 2027), advised by Sebastian (Basti) Scherer and Wenshan Wang. I received my B.S. (Honours) in Data Science from Hong Kong Baptist University in July 2026.

I first joined AirLab as a research assistant in April 2026. My work there has included foundation models for inertial perception — learning physically meaningful motion representations from raw IMU sequences, and building controllable generative models of inertial data to support robot state estimation.

Before moving into robotics, my research centered on medical image analysis: I developed segmentation-guided diffusion models for anatomy-preserving MRI generation (ICASSP 2026), deep-learning pipelines for material microstructure segmentation (AIPR 2025), and GAN-based super-resolution for diabetic retinopathy screening (ICDSE 2025).

Research interests: robot perception and state estimation, foundation models for sensing, generative models.

Portrait of Yi Zhao at Stanford University
Taken at Stanford, CA

News

Education

Carnegie Mellon University

Aug 2026 – Dec 2027

M.S. in Computer Vision, Robotics Institute

AirLab · Advised by Sebastian (Basti) Scherer and Wenshan Wang.

Hong Kong Baptist University

Sep 2022 – Jun 2026

B.S. (Honours) in Data Science

University of California, Berkeley

Jul 2024 – Aug 2024

Summer Session — Probability & Statistics, Finance

Experience

  1. AirLab, Robotics Institute, Carnegie Mellon University

    Apr 2026 – Present

    Master's Student · Advised by Sebastian (Basti) Scherer and Wenshan Wang

    Joined as a research assistant in April 2026; now a master's student in the MSCV program.

    Learning-based state estimation and inertial perception for field robotics: foundation models that learn 3-D motion representations from raw 6-axis IMU sequences across heterogeneous platforms (vehicles, quadrupeds, drones, humans), and controllable generative modeling of inertial time series for robot state estimation.

  2. Chinese Academy of Sciences

    Jan 2025 – Feb 2025

    Research Assistant

    Built an SRGAN-based super-resolution pipeline for diabetic retinopathy fundus images on the APTOS dataset — preprocessing, architecture design, perceptual-loss optimization, and two-stage adversarial training; the work led to a paper accepted at ICDSE 2025.

  3. China United Network Communications (China Unicom)

    Jun 2024 – Jul 2024

    Machine Learning Engineer Intern

    Analyzed 10k+ user records and built PyTorch forecasting models for provincial mobile-data demand (96% accuracy), engineering temporal and geographic features to support promotion strategies.

Publications

The Unexpired Plan: A Free Monitor for Accelerated Diffusion Policies

Yi Zhao, Sebastian Scherer

arXiv preprint · 2026

A training-free monitor for accelerated diffusion policies that compares actions with the unexecuted portion of a stored reference plan, with no additional reference forward pass for the check.

SegMaDiff: Segmentation-Guided Mamba Dynamic Diffusion Model for Anatomy-Preserving MRI Generation

Yi Zhao, Qirui Fan, Xiaohui Duan, Donglong Chen

IEEE ICASSP 2026

SegMaDiff injects tumor segmentation masks at every denoising step of a Mamba-based dynamic diffusion model, achieving state-of-the-art FID on BraTS-2023/2024 while preserving anatomical structure, and enabling mask-only restoration of corrupted MRI slices.

Automated Microstructure Segmentation in Titanium Alloys Using Deep Learning Techniques

Ruilang Wang, Yi Zhao, Ziqi Ye, Bowen Liu, Yucheng Li, Donglong Chen

AIPR 2025

★ Best Paper

A deep-learning pipeline (VGG U-Net and SegFormer) for TC4 titanium alloy metallographic images, reaching 90% pixel accuracy and 0.7 mIoU on 1024×1024 images with a customized IWMID + Otsu preprocessing workflow.

Super-Resolution Image Generation for Diabetic Retinopathy Detection by SRGAN

Yi Zhao

ICDSE 2025

An SRGAN pipeline with a VGG-19 perceptual module and two-stage adversarial training for fundus image super-resolution on the APTOS dataset, improving texture fidelity for downstream retinopathy screening.

Fun Fact