Chong He

Chong He (何翀)

PhD Candidate @ SFU
Machine Learning & AI
Robotics · Reachability
Human Prediction

About

- I am a PhD candidate specialized in Computing Science at Simon Fraser University, advised by Dr. Mo Chen (2023 – present). Check out the lab website: Mars Lab.
- My supervisory committee includes Dr. Sylvia Herbert and Dr. Xue Bin (Jason) Peng.
- I obtained my M.S. in Mechanical Engineering at University of California, San Diego, advised by Dr. Sylvia Herbert in The Safe Autonomous Systems Lab (2021 – 2023).
- I worked as a research intern at MaRobot AI in Vancouver on probabilistic human prediction with online intention inference (2025 – 2026).
- Previously, I worked as a research assistant in the Institute of Medical Robotics at Shanghai Jiaotong University (2020 – 2021).
- I received my B.Eng degree at Tongji University, China majoring in Mechanical Design, Manufacturing and Automation (2015 – 2020).
- I was also an exchange student at University of California, Berkeley (2019).
- My research interests lie in Machine Learning and Artificial Intelligence for robotics: human behavior prediction, inverse reinforcement learning, and safe autonomy (Hamilton-Jacobi reachability).

Skills

- Programming languages: Python, MATLAB, C/C++, and LaTeX.
- Packages & tools: JAX, HeteroCL, PyTorch, TensorFlow, NumPy, OpenCV, ROS/ROS2, and Gazebo.
- Languages: English and Mandarin.
- Love skiing and playing video games at leisure time.

Publications

- Minh Bui*, Chong He*, Jacob Bayless, Mo Chen, Rethinking Intention as Utility Function: Confidence-Based Prediction of Human Intentions for Social Robot Navigation, IEEE Transactions on Robotics (In preparation, 2026.)

- Minh Bui, Hanyang Hu, Chong He, Michael Lu, George Giovanis, Arrvindh Shriraman, Mo Chen, "Optimized_dp: An efficient, user-friendly library for optimal control and dynamic programming", Under review at ACM Transactions on Mathematical Software, 2025 Paper Code

- Chong He, Mugilan Mariappan, Keval Vora, Mo Chen "Threshold Strategy for Leaking Corner-Free Hamilton-Jacobi Reachability with Decomposed Computations", IEEE Conference on Decision and Control, 2025 Paper Code Project

- Frank J. Jiang*, Kaj Munhoz Arfvidsson*, Chong He, Mo Chen, Karl H. Johansson, "Guaranteed Completion of Complex Tasks via Temporal Logic Trees and Hamilton-Jacobi Reachability", IEEE Conference on Decision and Control, 2024 Paper Code

- Chong He*, Zheng Gong*, Mo Chen, Sylvia Herbert, "Efficient and Guaranteed Hamilton-Jacobi Reachability via Self-Contained Subsystem Decomposition and Admissible Control Sets", IEEE Control Systems Letter, 2023 Paper Code

- Xiaojie Ai, Anzhu Gao, Member, IEEE, Zecai Lin, Chong He, Weidong Chen, Member, IEEE “A Multi-Contact-Aided Continuum Manipulator With Anisotropic Shapes”, IEEE Robotics and Automation Letters, July 2021 IEEE

Patents

- Chong He, Mo Chen, Jacob Daniel Bayless, Michael Samuel Lu, Nhat Minh Bui, "Methods and Systems for Predicting One or More Future States of an Entity", PCT International Patent Application, filed Feb 2026 (pending)

Internship Experiences

MaRobot AI (Vancouver, British Columbia, Canada) — Research Intern, 04/2025 - 03/2026
- Formulated human intention prediction as utility (reward) inference with online Bayesian confidence updates
- Predicted distributions over pedestrians’ future states; deployed in real-world scenarios with ROS2 in Python
- Parallelized models in JAX; led to a co-first-authored manuscript and a PCT patent application

Shanghai Jiaotong University – Institute of Medical Robotics — Research Assistant, 06/2020 - 02/2021
- Examined the kinematics of a continuum manipulator using visual data
- Constructed experimental platforms for several projects
- Developed ROS software in C++
- Contributed to a paper published in IEEE Robotics and Automation Letters (2021)

Teaching Assistant Experiences

- CMPT 410/726: Machine Learning, Simon Fraser University, 2026 Fall

- CMPT 419: Robotic Autonomy, Simon Fraser University, 2026 Spring

- CMPT 410/726: Machine Learning, Simon Fraser University, 2024 Spring

Research

I. Human Trajectory Prediction in Human-Populated Environments SFU · 2025–present

How can a robot predict where people will walk in crowded, human-populated spaces — while understanding how people react to each other?

Self-driving cars and service robots operate in human-populated environments: busy streets, campuses, shopping malls, hospitals, and offices, surrounded by people. Working safely there depends on anticipating human motion: a vehicle yielding to a pedestrian about to step off the curb, or a service robot weaving through a crowded corridor. Two properties largely decide whether a trajectory prediction method works in these real crowds: whether it generalizes to new scenes, and whether it is socially aware, capturing how people influence one another. Yet they are rarely studied together.

My work

- Real-world deployment: at MaRobot AI, I formulated human intention prediction as utility (reward) inference with online Bayesian confidence updates, predicting distributions over pedestrians’ future states — deployed in real-world scenarios with ROS2 and parallelized in JAX.
- Intention as utility: confidence-based prediction of human intentions for social robot navigation T-RO paper (in preparation)
- Patent: Methods and Systems for Predicting One or More Future States of an Entity PCT patent (pending)

What's next

I aim to build human trajectory prediction that is both generalizable and socially aware — reliable enough for self-driving cars and service robots to work safely among people. Coming up:
- Confidence-based intention prediction: scaling to multiple utility-based intention functions and benchmarking against state-of-the-art methods.
- Survey paper: a survey on human trajectory prediction is on its way.
- Socially-aware prediction: new work on predicting multiple people together, capturing how they influence one another.

Trajectory PredictionSocial NavigationProbabilistic InferenceJAXROS2Python
II. Scalable Safety Guarantees: Fixing the Leaking Corner in HJ Reachability UCSD & SFU · 2022–present

How can we compute safety guarantees for high-dimensional robots — fast, and without giving up correctness?

Hamilton-Jacobi (HJ) reachability computes the set of states from which a robot can reach a goal or avoid failure, together with the optimal control policy. It comes with strong guarantees, but its cost grows exponentially with the state dimension (the "curse of dimensionality"). Decomposing a system into self-contained subsystems makes the computation tractable, but it can produce wrong results — errors that appear at the corners of the decomposed sets, which we call the leaking corner issue. My work identifies why this happens and how to fix it.

1. Admissible Control Sets

- Cause: the leaking corner comes from inexact decomposition of the target set and inconsistent control policies across the coupled subsystems.
- Fix: an admissible control set enforces consistent control actions across subsystems, giving exact reachable sets and optimal policies for decomposable targets.
- Beyond exact decomposition: conservative under-approximations for inexactly decomposed targets, which a local update in the full-dimensional space can refine to the exact result.

Exact versus inexact decomposition of a target set Reachable sets computed with admissible control compared with the true reachable set

Left: exact vs. inexact decomposition of the target set. Right: the backward reachable set from admissible control matches the true one, while plain subsystem decomposition does not.

Chong He*, Zheng Gong*, Mo Chen, Sylvia Herbert, "Efficient and Guaranteed Hamilton-Jacobi Reachability via Self-Contained Subsystem Decomposition and Admissible Control Sets", IEEE Control Systems Letters, 2023

L-CSS 2023 paper Code

Implementation: the admissible control set is implemented to guarantee the completion of complex tasks with temporal logic trees and HJ reachability:
Frank J. Jiang*, Kaj Munhoz Arfvidsson*, Chong He, Mo Chen, Karl H. Johansson, "Guaranteed Completion of Complex Tasks via Temporal Logic Trees and Hamilton-Jacobi Reachability", IEEE Conference on Decision and Control, 2024

CDC 2024 paper Code ReachabilityOptimal Control

2. Threshold Strategy

- Theory: a proof that the leaking corner issue occurs only at the corners of the decomposed sets.
- Algorithm: a local updating procedure that repairs just those corners, recovering the exact reachable set while keeping the speed of decomposed computation.

Illustration of the proof that leaking only happens at the corner Local updating process for the leaking corner

Left: why leaking can only happen at the corner. Right: the local updating process that recovers the exact reachable set.

Chong He, Mugilan Mariappan, Keval Vora, Mo Chen, "Threshold Strategy for Leaking Corner-Free Hamilton-Jacobi Reachability with Decomposed Computations", IEEE Conference on Decision and Control, 2025

CDC 2025 paper Code Project ReachabilityParallel Computation

3. Real-Time Safety for Arbitrarily Shaped Target Sets Ongoing

- Building on the Threshold Strategy: applies it to resolve leaking corners in decomposed HJ reachability.
- Any shape: combines simple-shape backward reachable tubes (BRTs) into targets and obstacles of arbitrary shape.
- Real time: solves only the leaking corners online, so BRTs are obtained fast enough for real-time robotics.

III. 3D Catheter Reconstruction from Image Sequences UCSD · 2022 IV. Teleoperation for Catheter Control SJTU · 2021

Awards

- Best Design Award, "Harting Cup" Undergraduate Science and Technology Innovation Competition, Tongji University, 2018

Contact

E-mail: chong_he@sfu.ca