DribbleAMP

Adversarial Motion Priors for Football Dribbling

Trained a generic humanoid and Unitree G1 to dribble a football using Adversarial Motion Priors for natural, human-like locomotion

Designed a velocity-conditioned ball reward and two-phase curriculum on top of MimicKit, trained with PPO in the Newton GPU simulator

Final Project for ETH Digital Humans Course

SmolVLA + Reasoning

Fine-tuned and deployed a SmolVLA policy on a real world SO-101 arm for pick and place tasks

Implemented a hierarchical System 1 + System 2 reasoning approach to enable OOD task completion

Final project for ETH Robot Learning Course with Oier Mees

Do Robots Dream of Better Actions?

World Action Models vs. VLAs on Cross-Embodiment Transfer and Learning from Heterogeneous Data

Empirically tested two claimed advantages of World Action Models (WAMs) over action-space VLAs: stronger cross-embodiment transfer and better learning from heterogeneous data

Benchmarked the MimicVideo video-action model against state-of-the-art VLAs (SmolVLA and Pi0.5) across two controlled experiments isolating each claim

Evaluated cross-embodiment transfer by adapting MimicVideo to three structurally different embodiments with limited play data, and compared homogeneous vs. heterogeneous fine-tuning data composition

Semester Project for ETH 3D Vision Course

LeWAM

Latent Space World Action Model

Designed and implemented a latent space world action model for multi-task robotic manipulation using flow matching

Joint video prediction and action generation in V-JEPA2 latent space using a single DiT backbone, inspired by DreamZero and LeWorldModel

Integrated as a LeRobot policy plugin for standardized training and real-world deployment on SO-101 arm

All datasets and models publicly available on Hugging Face

Zima

Autonomous Rover with end-to-end control

Engineered tracked rover and 5DOF robotic arm from CAD design to ROS2 control implementation

Trained and evaluated a behaviour cloning visuomotor policy (Resnet + MLP architecture) for real world navigation tasks

UBCO Marine Robotics

Software Lead

Led Marine Robotics software team competing at RoboSub 2024 and 2025. Ranked 11th internationally, 2nd in Canada

Designed modular autonomy and control systems using ROS2, Eigen, PCL, and PyTorch

Developed real-time semantic 3D mapping using YOLO object detection and stereo camera pointclouds

Implemented sensor fusion (DVL + IMU) for accurate navigation using PID and hierarchical state machines

Cellula Robotics

Software Intern

Developed ROS2 C++ and Python software for long-range AUV control and navigation systems

Migrated system-critical legacy ROS code to modern ROS2, developed CI/CD pipelines in GitLab

UBC

Research Assistant

Prototyped visualization solutions in Unity (C#) for large-scale curved displays

Integrated real-time motion capture (Motiv) and multi-system communication via Firebase

Developed mesh projection algorithms to handle geometric distortions on curved surfaces

Surveyed OOV visualization literature and proposed novel techniques adapted to large curved displays

Designed a within-subjects user study with NASA-TLX cognitive load measurement

Work done with Dr. Mohammad Khalad Hasan at UBC

ETH Zürich

Visiting Student, Exchange Program

Zürich, Switzerland

Coursework: Robot Learning, 3D Vision, Digital Humans (Motion Modeling), Philosophy of Language and Computation, German A1

University of British Columbia

Honours BSc, Major in Computer Science, Minor in Psychology

Kelowna, Canada

Coursework: Machine Learning (A+), Applied Linear Algebra (A+), Database Systems (A+), Image Processing (A+)

Achievements: 4.0 GPA / 93% Average, Dean's Scholar List, Deputy Vice-Chancellor Scholarship