CV

Professional CV covering robotics, computer vision, research, engineering experience, and technical projects.

Contact Information

Name Yue Xin (辛约)
Professional Title Ph.D. Student | Robot Learning, Embodied AI, and Intelligent Systems
Email xinyue4496@163.com
Location Beijing / Shanghai
Website https://iconssss.github.io

Professional Summary

Direct Ph.D. student at Tsinghua University with interdisciplinary training in physics, mathematics, measurement and control, and complex sensing systems. I combine rigorous modeling and experimental engineering with hands-on work in VLA fine-tuning, large-scale closed-loop evaluation, visuomotor policy learning, computer vision, 3D perception, ROS2 policy execution, and latency-aware robot systems. I am seeking robot learning and embodied AI research engineering roles that value both scientific discipline and end-to-end implementation.

Experience

  • 2026.06 - 2026.07

    China

    Industry Practice — Computer Vision Algorithms and Edge Deployment
    Goertek Inc.
    Developed and delivered a multi-person motion perception and identity recognition system for intelligent-device interaction, covering the full workflow from requirements and data collection to evaluation, GPU optimization, Unity integration, and Windows edge deployment.
    • Built a multi-person pose and motion-analysis pipeline with YOLO Pose and OpenCV, including ROI/slot-based perception and a causal temporal state machine for periodic-motion recognition.
    • Implemented temporal keypoint fusion, scale normalization, motion gating, and event-level manual annotation and matching for systematic evaluation.
    • Applied FP16 inference, temporal micro-batching, asynchronous video acquisition, and geometry-consistent preprocessing to achieve approximately 30 FPS video processing on an RTX 3050.
    • Evaluated a three-person historical test video with 145 ground-truth events, 142 predictions, and 141 matches, corresponding to an approximately 97.2% event match rate.
    • Built a local identity module with YuNet and SFace and established the binding workflow between student identity and visual slots.
    • Integrated heartbeat, identity, task-control, and real-time result messages with a Unity client through UDP/OSC, then verified delivery on a Windows RTX 3050 edge workstation.

Education

  • 2024 - Present

    Beijing, China

    Ph.D. Student (Direct Ph.D. Track)
    Department of Precision Instrument, Tsinghua University
    Precision Instrument
    • Researching complex inertial sensing systems through dynamic modeling, numerical simulation, error analysis, signal processing, parameter calibration, and experimental platform integration.
    • Developed experience in reasoning about sensing uncertainty, system identification and debugging, and hardware-software interaction.
  • 2020 - 2024

    Beijing, China

    Bachelor's Degree, Dual-degree Program
    Weiyang College, Tsinghua University
    Mathematical and Physical Basic Sciences; Measurement, Control Technology and Instruments
    • Completed interdisciplinary training across physics, mathematics, numerical analysis, control, measurement, and instrumentation.
    • Built a rigorous foundation for cross-domain research and engineering problem solving.

Skills

Languages: Python, C++, MATLAB
Machine Learning and Vision: PyTorch, OpenCV, YOLO Pose, YuNet, SFace, computer vision, deep learning, 3D perception
Robot Learning: ACT, imitation learning, Behavioral Cloning, Diffusion Policy, action chunking, closed-loop rollout evaluation, LeRobot, SmolVLA
Robotics: ROS2, MoveIt, ros2_control, KDL inverse kinematics, robot manipulation, motion planning
Simulation: ManiSkill, MuJoCo, LIBERO, Panda robot simulation
ML and Runtime Systems: CUDA environment debugging, multi-GPU DDP, DeepSpeed ZeRO, checkpoint management, asynchronous policy execution, policy latency analysis, evaluation RNG control, model/dataset contract checking
Engineering Tools: Linux, Windows, WSL2, Git/GitHub, Docker, SSH, UDP/OSC, Unity integration
Modeling, Control, and Experimentation: dynamic system modeling, numerical simulation, error propagation, parameter estimation and calibration, signal processing, PID control, hardware-software integration, system troubleshooting

Languages

Chinese · Native speaker
English · CET-6 596 · Able to read research papers and communicate in technical contexts

Projects

  • SmolVLA Fine-Tuning and 400-Episode Closed-Loop Evaluation

    Completed an auditable, single-RTX-4090 SmolVLA adaptation workflow from pinned pretrained-policy initialization to a frozen 100K-update checkpoint and four-suite LIBERO evaluation.

    • Fine-tuned for 100,000 optimizer updates at effective batch 64 (6.4M sample presentations), with 99.88M / 450.05M parameters trainable and the vision encoder frozen; preserved frozen configs, hashes, logs, and rollout evidence.
    • {“Evaluated 40 tasks over 400 fixed-init closed-loop episodes, reaching 71.5% endpoint success overall”=>”82% Spatial, 78% Object, 80% Goal, and 46% Long.”}
    • Resolved the checkpoint-to-LIBERO 3-camera/6D/6D versus 2-camera/8D/7D runtime-contract mismatch; verified strict offline reload with zero missing, unexpected, or shape-mismatched keys.
  • Latency-Aware Runtime for Closed-Loop Robot Policies

    Designed an asynchronous learned-policy runtime that tracks action provenance and controls scheduler-induced staleness through explicit synchronous, FIFO, and latest-only semantics.

    • Decomposed action age into observation wait, inference latency, and post-inference wait, with per-action timestamp and queue-depth instrumentation.
    • On MuJoCo Reacher at 100 ms policy latency, LATEST reduced p95 action age from 1.24 s to 150 ms and improved success from 20% to 100% while preserving approximately 20 Hz control.
    • Validated the mechanism on both a learned action-chunk benchmark and an independent MuJoCo environment with a fixed learned SAC policy.
  • Metric 3D Representations for Low-Data Visuomotor Control

    Built a calibrated RGB-D to world-frame point-cloud pipeline and a compact PointNet policy, then mapped sample efficiency, viewpoint robustness, occlusion recovery, and calibration failure boundaries.

    • Achieved 95% closed-loop success with 50 demonstrations and 100% with 250 demonstrations on a contact-free MuJoCo Panda reach task.
    • Retained 96.7–100% success across 0–45° camera motion by reconstructing observations into a common world frame.
    • Restored strong-occlusion performance from 3.3% to 100% through two-view fusion without policy retraining; quantified collapse under depth and extrinsic error.

Research and Engineering Focus

  • Visuomotor policy learning for manipulation, including imitation learning, action chunking, replanning, generalization, and closed-loop evaluation.
  • VLA training and evaluation systems, with emphasis on model/data/action contracts, cross-embodiment transfer, checkpoint comparison, and failure diagnosis.
  • Metric 3D and multi-view perception for robot interaction, including calibration sensitivity and robustness under distribution shift.
  • Latency-aware robot policy runtimes, asynchronous inference semantics, ROS2 execution, motion planning, and reliable control.
  • Reproducible embodied AI experimentation through controlled baselines, ablations, multi-seed evaluation, provenance tracking, and honest negative-result analysis.

Currently Exploring

  • VLA and robot foundation model adaptation for manipulation and cross-embodiment transfer.
  • Recovery-aware imitation learning, multimodal behavior, and flow-based robot policies.
  • OpenPI and π0-family training, evaluation, and deployment workflows.

Professional Strengths

  • Rigorous foundation in physics, mathematics, control, and system modeling, supported by Tsinghua dual-degree training.
  • Ability to move between model formulation, software implementation, evaluation, and experimental or edge-system debugging.
  • Hands-on experience across research prototypes, simulation benchmarks, GPU training systems, and delivered computer-vision applications.
  • Transferable systems perspective connecting sensing uncertainty, control, policy learning, and embodied execution.

Honors and Awards

  •
    National High School Mathematics League (Shanghai Division), Second Prize
    Competition Award
  •
    National High School Physics Competition (Shanghai Division), Second Prize
    Competition Award