Carnegie Mellon · Monash
I build intelligent systems that perceive physical properties and causal rules from visual data, toward cognitive digital twins.
Recovering geometry and motion of dynamic scenes from monocular video in real time.
Estimating mass, friction, and contact from observation, so a reconstruction can be simulated forward rather than merely replayed.
Perception for loco-manipulation: what a robot must infer about a scene before it makes contact.
A spatially-factorized motion field decouples geometry from non-rigid dynamics through a learnable per-point kinematic subspace. Renders above 163 FPS, stores in ~11 MB, trains in ~18 minutes, and needs no SfM initialization.
A dual-network architecture separates fine-grained local deformation from global motion under optical-flow guidance, resolving the usual trade-off between detail preservation and temporal coherence in dynamic scenes.
A mixture-of-experts framework orchestrates open-source LLMs into executable Python for compositional visual tasks, reaching state-of-the-art accuracy with no training or fine-tuning.
A self-attention deep temporal convolutional network predicts non-linear tremor signals far enough ahead for closed-loop active suppression.
Robotics Institute — AirLab, Carnegie Mellon University
Host: Prof. Sebastian Scherer
Monash University
4D reconstruction, simulation, and robotics
Monash University
Intel Corporation
Design Enablement Group — DTCO, APR team
ViTrox Corporation Berhad
FIT5215Deep LearningPostgraduateFIT5222Planning and Automated ReasoningPostgraduateFIT5217Natural Language ProcessingPostgraduateFIT5216Modelling Discrete Optimization ProblemsPostgraduateFIT3181Deep LearningUndergraduate