Portrait of Yecheng (Geoffrey) WANG

CV / GitHub / LinkedIn

mn232yw@leeds.ac.uk

Yecheng (Geoffrey) WANG 「王烨澄」

Applying to PhD / MS programs in robotics · Fall 2027

BEng (Hons) Mechanical Engineering @ University of Leeds BEng Mechanical Design, Manufacturing and Automation @ Southwest Jiaotong University (SWJTU) Expected Graduation: Jul 2027 (2+2 Joint Program) Final-Year Project: Learning-based Control for a Rehabilitation Robot @ Rehab45 Lab, supervised by Dr. Justin Gallagher

Standing at the convergence of Embodied AI and Robots, I am determined to make the most advanced technology affordable and accessible.

Goal
Surgical Robots that are precise enough to trust and simple enough to afford.
Research Interests
Robot Learning · Robot Modeling and Control · Learning-based Autonomy · Physically Grounded Motion Prediction and Planning.
Current Question
How few DoF does a flexible endoscopic instrument need?
How can we close the gap between commanded, desired and actual motion?

Research

Surgical Robotics & Instrumentation Lab (SRIL), CUHK, Hong Kong SAR Jun 2026 – Present

PI: Prof. Shing Shin CHENG · Mentor: Dr. Jun HUO

Poster: From Millimetres to Microns (CUHK SURP 2026)

From Millimetres to Microns: Hysteresis Compensation of a Cable-Driven Continuum Surgical Robot

Jun 2026 – Aug 2026 · CUHK Summer Undergraduate Research Programme

Cable hysteresis was the dominant error source; compensating it cut tip error 3.0× (234 → 78 µm), and an iterative learning layer brought it to ~15 µm.

I rebuilt a 14-notch tendon-driven continuum surgical robot with stereo tip tracking and cable-tension sensing, and identified path-dependent cable hysteresis as the dominant error source. I then designed a data-driven hysteresis compensator and a P-type iterative learning controller for the 0.25 s vision latency, and showed that tip loading widens the hysteresis loop, contradicting a friction-only assumption. The compensator was validated under a preregistered protocol on a sealed multi-target benchmark; the learning layer then approached the platform’s repeatability limit.

poster / project page

Task-Conditioned Kinematic Coverage of Reduced-DoF Continuum Robots for Endoscopic Surgery

Aug 2026 – Present

Can a fixed-structure 4-DoF continuum arm execute each surgical gesture within position and tool-axis tolerances — and if not, why not?

I built a kinematic simulation that tests each gesture against these tolerances, attributes failures to reach or position–axis coupling, and benchmarks four comparators. Next, I plan to replace synthetic primitives with instrument trajectories from open dVRK datasets, identify the arm’s bending structure from EM tip tracking, and validate per-gesture predictions in teleoperation on a dual-arm flexible endoscopic platform.

Physically Grounded Counterfactual World Models for Explainable Motion Prediction and Planning

Dec 2025 – Present · KineWorld (AAAI 2027, advanced to Phase 2) · PHLEX (ICLR 2027, under review)

Explain a frozen predictor’s trajectories with differentiable motion mechanisms, then re-run those explanations as physically grounded counterfactual futures.

KineWorld overview: standard multimodal predictors show mode collapse or kinematically infeasible trajectories; a frozen predictor with KineWorld yields diverse, feasible ones
KineWorld

KineWorld is a pluggable world model that decodes a frozen predictor’s trajectories against a library of differentiable motion mechanisms, giving interpretable behavior attribution and counterfactual rollouts; on UniAD/nuScenes it lowered maximum yaw rate from 0.986 to 0.640 rad/s and raised mechanism self-consistency from 8.45% to 57.27%, with gains in comfort, safety and planning accuracy.

PHLEX overview (excerpt): a black-box predictor’s trajectories explained through kinematic, control-law and scene-reference layers
PHLEX

PHLEX extends the library into a predictor-agnostic framework that explains predicted trajectories through layered kinematic models, control laws and scene references, and re-runs varied explanations as counterfactual futures for ego planning. On three frozen planners (UniAD, VAD, SparseDrive) it cut collision rate by up to half and yaw-rate violations by 1.9–6.7 pp, and in agent forecasting (UniAD/MTR/QCNet on nuScenes/Argoverse 2) it reduced kinematic violations by >80% for ≤0.05 m minFDE.

Co-researcher, with a PhD student at Purdue University

Publications

Honors

Selected Projects

More projects
  • Helicopter Altitude Control Mar 2025 – Apr 2025
    Implemented and tuned a PID altitude controller for a tethered single-axis helicopter rig, run remotely on the University of Leeds ReLOAD platform: take-off, timed hover within ±2 cm of the target height, and landing. video / code
  • Bridge Load Analysis, Simulation and Manufacture Mar 2025 – Apr 2025
    Modeled a 2D truss bridge from engineering statics, solved the 24-equation system in MATLAB for load simulation, and built a MATLAB GUI that predicts vulnerable members and maximum load capacity. GUI screenshot / truss model / code
  • YOLO-Based Automatic-Checkout Cart Apr 2024 – Apr 2025
    Built a Raspberry Pi 5 cart prototype fusing a fisheye camera and a load cell, with a 2-DoF pan-tilt laser pointer that flags unrecognized items; a YOLOv8n with Multi-Scale Dilated Attention reached 92.4% mAP@0.5 at <1 s/item on the edge device. demo video / slides

Internship