Angus Fung

I completed my Ph.D at the University of Toronto, Robotics Institute, advised by Goldie Nejat, where I work on robot perception and control using self-supervised learning and generative AI.

I am currently building Syncere with Aaron Tan to bring consumer robots into society.

Previously, I worked on learning algorithms at the Vector Institute where I was advised by Jimmy Ba.

1. Music: I am active as a church organist having held positions at the Metropolitan United Church (under Dr. Patricia Wright) and St. Michael's Cathedral Basilica. I received my ARCT Diploma in Piano and Organ Performance in 2013 at the Royal Conservatory of Music.
2. Medicine: I build AI models to tackle open questions in neuroscience and neuro-ophthalmology with Dr. Anthony Lang and Dr. Edward Margolin.
   
3. Startups: In the past, I co-founded Scholarply and ONE800 with Aaron Tan, leveraging LLMs for scholarship applications and personal companionship.
4. Entertainment: I worked with 2x Grammy Award recipient Sean Leon to build AI technology for their God's Algorithm Project.

Updated: 12/24

Email  /  Google Scholar  /  Github

We present a novel people tracking architecture for mobile service robots using conditional latent diffusion models, which we name Latent Diffusion Track (LDTrack), to solve the robotic problem of tracking multiple dynamic people under intraclass variations.

We introduce X-Nav, a novel framework for cross-embodiment navigation where a single unified policy can be deployed across various embodiments for both wheeled and quadrupedal robots.

We present MLLM-Search, a novel multimodal language model approach to address the robotic person search problem under event-driven scenarios with incomplete or unavailable user schedules. Our method introduces zero-shot person search using language models for spatial reasoning, a novel visual prompting method generating topological graphs with semantic labels, and an MLLM-based search planner combining region and waypoint planning through our spatial chain-of-thought (SCoT) prompting method.

We introduce a novel Hand-drawn Map Navigation (HAMNav) architecture that leverages pre-trained vision language models for robot navigation across diverse environments, hand-drawing styles, and robot embodiments, even in the presence of map inaccuracies.

We present Finder, a novel approach to the multi-object search problem that leverages vision language models (VLMs) to efficiently locate multiple objects in diverse unknown environments. Our method combines semantic mapping with spatio-probabilistic reasoning and adaptive planning, improving object recognition and scene understanding through VLMs.

We present a novel multimodal person detection architecture to address the mobile robot problem of person detection under intraclass variations (e.g. partial occlusion, varying illumination, pose deformation) by introducing our Temporal Invariant Multimodal Contrastive Learning (TimCLR) method.

We present the novel Weakly Supervised Mask Data Distillation architecture for autonomously generating pseudo segmentation labels.

We present a novel multi-robot person search system to generate search plans for multi-robot teams to find multiple dynamic users before a deadline.

We present a method to calibrate the time-varying extrinsic transformation between any number of cameras and achieves measurement excitation over the entire configuration space of the mechanism resulting in a more accurate calibration.

We investigate the first use of deep networks for victim identification in Urban Search and Rescue, for cases of partial occlusions and varying illumination, on a RGB-D dataset obtained by a mobile robot navigating cluttered USAR-like environments.