Projects

Dec 2025 Master’s Thesis · RSL, ETH Zürich

Multi-Domain Motion Embedding

Expressive Real-Time Mimicry for Legged Robots

Prior motion imitation methods rely on explicit, morphology-specific retargeting and fail to generalise to unseen motions because they do not capture the inherent periodic and aperiodic structure of natural movement. MDME addresses this with a dual-encoding architecture that combines a Variational Autoencoder for unstructured features with a discrete wavelet transform (DWT) encoder for periodic structure, producing a rich latent representation that conditions a robot policy directly on raw reference motions — no retargeting needed. Trained jointly on human (AMASS) and dog (SIGGRAPH) motion-capture datasets, the framework demonstrates zero-shot sim-to-real deployment on the Fourier N1 humanoid and ANYmal D quadruped, outperforming prior approaches in reconstruction fidelity and generalising to novel unseen motion styles in real time.

arXiv PDF

Sep 2023 Semester Thesis · RSL, ETH Zürich

Small-Scale Quadruped Depth-Based Exteroception

Computer Vision-Based RL for Single-Camera Obstacle Traversal

Small quadruped robots lack the proprioceptive richness of larger platforms, making robust obstacle traversal challenging without expensive sensor suites. This project developed a depth-camera RL controller trained entirely in IsaacSim with a novel stereo-IR noise simulation model to close the sim-to-real gap. The resulting policy traverses obstacles up to 15 cm tall and deploys zero-shot to hardware via an Oak-D Pro camera, with no fine-tuning on real data.

Report

Jan 2023 Course Project · ETH Zürich

Mixed Reality Floorball Goalie

Real-Time Attack Space Visualisation for Goalies

Goalies struggle to read fast-developing attack scenarios because threat space is abstract and hard to infer under pressure. This mixed reality application tracks the goal posts, player positions, and ball in real time to render the opponent’s attack cone as an AR overlay, giving goalies an immediate spatial reference. Built as the final project for the Mixed Reality course at ETH Zürich.

Report

May 2022 Bachelor’s Capstone · Georgia Tech — Sponsored by General Motors

Hybrid Motor Disengagement System

Electromagnetically Actuated Dog Clutch for Hybrid Electric Vehicles

In hybrid vehicles the electric motor must disengage from the transmission at high speeds to prevent motor overspeed failure, yet existing solutions are bulky and difficult to manufacture. Sponsored by General Motors, Team Matilda designed a compact electromagnetically actuated dog clutch: compression springs keep the clutch naturally engaged while energising an electromagnet pulls an armature to axially retract the outer dog clutch plate, decoupling motor from drivetrain in milliseconds. Rotary shaft encoders communicate speed over CAN to the ECU, enabling synchronisation before re-engagement. FEA on critical components, an FMEA, and material selection (low alloy steel for yield strength and machinability) validated the design; a 3D-printed proof-of-concept prototype demonstrated the engage / disengage mechanism. Bulk unit cost at 300 000 units came to $241 — within the $300 target.

Report