Tao Zhong

I'm a Ph.D. student at Princeton University, supervised by Prof. Christine Allen-Blanchette. I received my B.A.Sc. in Engineering Science from the University of Toronto, where I worked with Prof. Animesh Garg. Previously, I interned at Noah's Ark Lab Canada with Prof. Yang Wang on computer vision and domain generalization, and worked with Prof. Huihuan Qian at AIRS and CUHK(SZ) on marine robotics.

Feel free to check out my CV or send me an Email if you want to connect.

Email  /  CV  /  Scholar  /  LinkedIn  /  Github

Research Interests

My research focuses on Physics- and Geometry-Aware Deep Learning. I develop novel machine learning and generative techniques that exploit the mathematical structure of the physical world, including symmetries, topology, and physical laws, to solve complex problems in:

• Scientific Discovery: Designing structure-preserving representations and physics-aware frameworks to model complex dynamics and solve high-dimensional inverse problems.
• Robot Learning: Leveraging geometric priors and generative models to enable sample-efficient policy learning and robust decision-making for manipulation and control.
• Some of my early work also focused on domain adaptation and generalization, building models that can robustly adapt to new tasks and domains.

Publications

We introduce a geometric neural operator endowed with an algebraic-level inductive bias to explicitly preserve global topological structures in physical systems.

We propose NeFTY, a framework that enables accurate 3D thermal tomography by combining continuous neural fields with a differentiable physics solver to recover subsurface defects from surface temperature data.

We introduce LEGO, a symmetry-aware graph neural network framework that enables sample-efficient, scalable, and generalizable swarm robot control across diverse team sizes and environments.

Grasp2Grasp enables simulation-free, vision-based translation of dexterous grasps across robot hands using Schrödinger Bridges with physics-aware costs for stable, functionally aligned grasps.

GAGrasp uses a geometric algebra diffusion model to generate robust, physically plausible dexterous grasps that are naturally equivariant to an object's pose.

We introduce VDPG, a method that adapts large models to new visual domains at test-time using only a few unlabeled images to generate a domain-specific prompt that guides the model's features.

This paper presents Fast-Grasp'D, a differentiable simulator that rapidly generates Grasp'D-1M, a large dataset of stable, contact-rich, multi-finger grasps for robotic learning.

We propose Meta-DMoE, a framework that adapts to domain shifts by using a meta-learned aggregator to distill knowledge from specialized expert models into a student network for fast test-time adaptation.

Education

Miscellanea