avatar

Xiaoyi Lin

Robotics & Autonomous Systems Research Engineer
Ambarella Inc.
linxiaoyi1108@gmail.com


About Me

Hi, I am Elsie (Xiaoyi) Lin, a Machine Learning Engineer at Ambarella working on autonomous driving perception and robotics systems. I received my M.S. degree in Computer Engineering from New York University, where I conducted research in the Control and Networks Lab (CAN Lab) and AI4CE Lab.

My work focuses on deep learning-based perception systems, multi-modal sensor fusion, trajectory prediction, and real-time AI deployment for autonomous systems. At Ambarella, I develop production-grade radar perception DNN pipelines using high-resolution imaging radar and Transformer-based architectures for traffic actor detection, tracking, free-space reasoning, and autonomous driving behavior understanding. I also work on Quantization Aware Training (QAT), model compression, and embedded deployment optimization for real-time inference on resource-constrained autonomous driving platforms.

Previously, I worked at Amazon Robotics on motion planning and machine-learning-based collision checking systems for robotic manipulation and autonomous navigation. My research experience spans autonomous driving, reinforcement learning-based motion planning, SLAM, mapping/localization, diffusion-based planning systems, and large-scale simulation/testing infrastructure.

I am actively seeking PhD opportunities in Robotics and Embodied AI.

Technical Skills

Work Experience

  1. Machine Learning Engineer

    Oculii (An Ambarella Company), Beavercreek, OH

    • Developed production-grade autonomous driving systems using high-resolution imaging radar and deep learning pipelines.
    • Improved traffic actor classification accuracy from 60% to 82% using multi-modal Transformer-based perception systems.
    • Conducted research on Quantization Aware Training (QAT), model compression, and real-time embedded deployment optimization.
    • Developed Occupancy Grid Mapping and free-space reasoning systems for navigation-aware autonomous driving applications.

  2. C++ SDE Intern

    Amazon Robotics, North Reading, MA

    • Adopted and implemented an up-to-date machine learning-based algorithm for collision detection using C++.
    • Effectively integrated the algorithm with motion planning code base; reduced 5.4% overall computational expenses for each planning, improving collision checking run-time from ~1ms to ~100ns.
    • Developed test strategy and conducted simulations on the Nvidia PhysX platform and AWS S3.

  3. R&D Intern

    Volkswagen, Shanghai, China

    • Improved a machine learning-based driver fitness determination system using Python. Resolved the overfitting problem by decreasing the sensitivity of RNN, reducing the error rate of blinking detection by 12%;
    • Adopted CI/CD to deploy the system onto the flagship model VW Tiguan using C++ and AUTOSAR

Patterns

Publications

  1. ITSC
    Learning-Based State Estimation for Automated Lane-Changing
    Won Yong Ha, Sayan Chakraborty, Xiaoyi Lin, Kaan Ozbay, Zhong-Ping Jiang
    International Conference on Intelligent Transportation Systems (ITSC), 2024.
    PDF Code Project Page

  2. CVPR
    A Compacted Structure for Cross-domain learning on Monocular Depth and Flow Estimation
    Yu Chen, Xu Cao, Xiaoyi Lin, Baoru Huang, Xiao-Yun Zhou, Jian-Qing Zheng, Guang-Zhong Yang
    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020.
    PDF Code Project Page

Researchs

  1. Advanced perception for learning-based auto lane-changing
    Control and Networks Lab, NYU
    • Leveraged perception for a learning-based auto lane-changing project using RGBD camera and LiDAR;
    • Developing an end-to-end motion planner with diffusion from learning model, enhanced the safety and efficiency of the vehicle under low-representative situations.

  2. Mapping/Localization/Navigation R&D for Self-Driving
    AI4CE Lab, NYU
    • Optimized a redundant mapping process by utilizing openVSLAM system, kept its robustness on Colmap but reduced offline processing time and supported more features than the current ORB filter.

  3. 3D Reconstruction Based on Depth-Optical Fusion DNN
    Oxford University, UK
    • Constructed an artificial neural network by PyTorch; completed the fusion of optical flow, depth and semantic information; implemented moving average method to optimize the framework;
    • Completed the high-quality reconstruction of three-dimensional images.

  4. Automatic Venipuncture Robot Based on Deep Learning
    Tongji University, Shanghai, China
    • Designed the compact supporting unit, positioning unit, puncturing unit, and imaging unit;
    • Designed a semi-supervised method for vein segmentation - Semi-ResNeXt-Unet, which can determine the depth of a vein in ultrasound images and hence navigate the puncture of VeniBot;
    • Collected ultrasound images of 400 patients, including 100 manually labeled patients and 300 unlabeled patients, to validate the semi-ResNeXt-Unet network.