Aaron Hao Tan

I just finished my PhD, advised by Dr. Goldie Nejat, at the University of Toronto, where I work on giving robots the ability to anticipate the world via deep MARL, diffusion, and language models.

I am an incoming postdoctoral researcher at Stanford (Spring 2025). Previously, I was nominated as a PhD Apple Scholar in AI/ML.

Updated: 10/24

Email  /  Twitter  /  CV  /  Resume  /  Google Scholar

We present a novel robot exploration map prediction method called Confidence-Aware Contrastive Conditional Consistency Model (4CNet), to predict (foresee) unknown spatial configurations in unknown unstructured multi- robot environments with irregularly shaped obstacles.

We introduce OLiVia-Nav, an online lifelong vision language architecture for mobile robot social navigation. By leveraging large vision-language models (VLMs) and a novel distillation process called SC-CLIP, OLiVia-Nav efficiently encodes social and environmental contexts, adapting to dynamic human environments.

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 propose NavFormer, a novel end-to-end DL architecture consisting of a dual-visual encoder module and a transformer-based navigation network to address for the first time the problem of TDN in unknown and dynamic environments.

The first Macro Action Decentralized Exploration Network (MADE-Net) using multi-agent deep reinforcement learning to address the challenges of communication dropouts during multi-robot exploration in unseen, unstructured, and cluttered environments.

The first extensive investigation of mobile robot inference problems in unknown environments with limited sensor and communication range and propose a new taxonomy to classify the different environment and task inference methods for single- and multi-robot systems.

The development of a novel sim-to-real pipeline for a mobile robot to effectively learn how to navigate real-world 3D rough terrain environments.

A 1:6 scale, multi-wheeled mobile robotic platform with independent suspension, steering and actuation for off-terrain operations.

These are works that I have contributed to during my free time.

I was acknowledged in the development of a smart clothing system using low-cost, energy-efficient strain sensors to measure elbow and shoulder joint angles. A novel neural network architecture was used to enhance the accuracy of the sensor signal mapping.

We propose a novel two-step deep learning (DL) method based on our extended ResNet-50 model. It can detect and classify whether face masks are missing, are worn correctly or incorrectly, or the face is covered by other means (e.g., a hand or hair).

We propose a novel process for a 6-DOF static one-armed robotic manipulator to address the challenges of pillow placement on a small-scale bed.

The first application of deep learning to the classification of homonymous defects on automated perimetry.

A competition designed to answer: how can robots respond and interact in an ethical manner when delivery objects within a domestic setting? Placed First.

Implemented image and position-based visual servoing on a Clearpath Husky and UR5 Manipulator.

• Accelerating the scholarship application process via LLM agents to help students secure funding while focusing on their studies.

• Selected by Microsoft Startup Hub Program, receiving grants worth $150k.

• An all-in-one service built in to iMessage aimed at lowering the barrier of entry for LLMs and Generative AI.