hjsuh at mit dot edu
CV | Google scholar | Github
Latest CV Update: 2024-01-03

Hyung Ju Terry Suh

Ph.D. candidate at MIT CSAIL
Advisor: Prof. Russ Tedrake at the Robot Locomotion Group.

Research Interest

My research interest lies in enabling robots with human-like dexterity and intelligence in manipulation. Through this technology I aim to broaden the spectrum of capabilities we have on automating physical tasks in the real world. To achieve this capability, I am searching for elegant solutions for perception, planning, and control through contact-rich dynamics that are frequent in real-world manipulation. In search of these solutions, I am broadly interested in bringing tools from rigorous model-based / mathematical disciplines (simulation, control, optimization, statistics) to understand, interpret, and improve recent empirical advances in machine learning (reinforcement learning, computer vision).

Background

Before joining MIT, I was an undergrad at Caltech (B.S. in Mech E and CS), with thesis advisors Joel Burdick and Aaron Ames. I also spent summers in Toyota Research Institute (TRI), and NASA JPL (Jet Propulsion Laboratory). Even before then, I was a sergeant in the Korean army, and graduated from Korean Minjok Leadership Academy.

Recent Events

Research

Fighting Uncertainty with Gradients: Offline Reinforcement Learning via Diffusion Score Matching
Offline optimization paradigms such as offline RL and imitation learning are highly sensitive to model bias, and need a scalable way to control the degree of uncertainty. We investigate distance-to-data as a candidate, and show that its gradients can be estimated via score-matching, which can be directly used in policy search. Combining these insights, we propose Score-Guided Planning (SGP), and show that it scales better than zeroth-order methods, and overcomes spurious local minina of ensembles due to statistical estimation.
  • H.J. Terry Suh, Glen Chou, Hongkai Dai, Lujie Yang, Abhishek Gupta, Russ Tedrake
    Fighting Uncertainty with Gradients: Offline Reinforcement Learning via Diffusion Score Matching [pdf][video]
    2023 Conference on Robot Learning (CoRL)
Global Planning for Contact-Rich Manipulation via Local Smoothing of Quasidynamic Contact Models
The classical obstacles for planning through contact has been associated with explosion of contact modes, and the non-convexity of the planning problem. We show that when we abstract contact modes by smoothing, classical algorithms such as RRT that were effective in tackling non-convexity can be used to effectively solve difficult tasks in highly contact-rich manipulation.
  • Tao Pang*, H.J. Terry Suh*, Lujie Yang, Russ Tedrake,
    Global Planning for Contact-Rich Manipulation via Local Smoothing of Quasidynamic Contact Models [pdf][video]
    2023 Transactions of Robotics (T-RO)
Do Differentiable Simulators Give Better Policy Gradients?
Differentiable simulation promises faster convergence for policy search by replacing typical zero-order policy gradients with first-order ones. Contrary to this belief, we find pathologies where using the first-order gradient actually hurts performance compared to the zero-order one. We propose a robust interpolation scheme between the two gradients that can alleviate the discovered pathologies. 
  • H.J. Terry Suh, Max Simchowitz, Kaiqing Zhang, Russ Tedrake,
    Do Differentiable Simulators Give Better Policy Gradients?
     [pdf][talk]
    2022 International Conference on Machine Learning (ICML), Outstanding Paper Award
  • H.J. Terry Suh, Max Simchowitz, Kaiqing Zhang, Tao Pang, Russ Tedrake,
    Pathologies and Challenges of Using Differentiable Simulators in Policy Optimization for Contact-Rich Manipulation
     [pdf]
    ICRA 2022 Workshop: Reinforcement Learning for Contact-Rich Manipulation
Bundled Gradients via Randomized Smoothing
How can we understand the success of RL for contact-rich manipulation? We focus on the stochastic formulation of RL approaches, and show that by abstracting different contact modes, the process and sampling and averaging leads to randomized smoothing of the contact dynamics. We show that this alleviates problems of applying gradient-based approaches for non-smooth dynamics, and utilize stochastic gradients for planning by modifying the exact linearizations in iLQR. 
  • H.J. Terry Suh*, Tao Pang*, Russ Tedrake,
    Bundled Gradients through Contact via Randomized Smoothing
     [pdf][video]
    2022 Robotics and Automation Letters (RA-L), Presented at 2022 ICRA
SEED: Series Elastic End-Effectors in 6D For Visuotactile Tool Use
Can we see visuotactile sensing as not only a new mechanism for sensing, but an opportunity to fundamentally change our hardware? By drawing an analogy between visuotactile sensors and series elastic actuators, we achieve both a generalization of series elastic actuators to 6D, and an abstraction of visuotactile sensors that is useful for force control. 
  • H.J. Terry Suh, Naveen Kuppuswamy, Tao Pang, Alex Alspach, Paul Mitiguy, Russ Tedrake,
    SEED: Series Elastic End-Effectors in 6D For Visuotactile Tool Use
     [pdf][video]
    2022 International Conference on Intelligent Robots and Systems (IROS)
  • H.J. Terry Suh, Naveen Kuppuswamy, Tao Pang, Alex Alspach, Paul Mitiguy, Russ Tedrake,
    SEED: Series Elastic End-Effectors in 6D For Compliant Visuotactile Tool Use
     [pdf]
    2022 ICRA Workshop: Compliant Robot Manipulation: Challenges and New Opportunities. 
  • R. Gordon, Soft Robots that Grip with the Right Amount of Force, MIT News [link
The Surprising Effectiveness of Linear Models for Visual Foresight in Object Pile Manipulation
How can robots push piles of small objects into a desired target set using visual feedback? This problem breaks a lot of conventional representations (the notion of Lagrangian states) that are used to model objects, and we are forced to reason in the space of observations. We show that surprisingly, a linear model outperforms deep model in predicting evolution of pixels for this problem.
  • H.J. Terry Suh, Russ Tedrake,
    The Surprising Effectiveness of Linear Models for Visual Foresight in Object Pile Manipulation
     [pdf][video]
    2020 International Workshop on the Algorithmic Foundations of Robotics (WAFR)
  • H.J. Terry. Suh,
    Predictive Models for Visuomotor Feedback Control in Object Pile Manipulation [pdf]
    MIT Thesis, S.M. degree
Hybrid Locomotion
Can we equip robots with multiple modalities of locomotion? A flying robot can fly anywhere, but choose to drive when it wants to save energy. These capabilities may be very important for robots that need to operate in rough terrain for extended period of time.
  • H.J. Terry Suh, Xiaobin Xiong, Andrew Singletary, Aaron Ames, Joel Burdick,
    Energy-Efficient Motion Planning for Multi-Modal Hybrid Locomotion [pdf][video]
    2020 International Conference on Intelligent Robots and Systems (IROS)
  • H.J. Terry Suh, Design and Planning of Flying-Driving Hybrid Robot [pdf][video]
    Bachelor’s Thesis, California Institute of Technology, 2019
Semantic Task Planning with Temporal Logic Synthesis
Someday, we wish robots to listen to natural language commands and plan the task on their own. This work assumes that the natural language can be converted to temporal logic, and uses temporal logic synthesis to convert that logic into a finite-automata which acts as a high-level reactive controller.
  • H.J. Terry Suh,
    Semantic Task Planning in Household Environments using Temporal Logic Synthesis [pdf]
    Self-hosted, 2018.
Vehicle Telelop with Impedance Controlled Manipulator
I have been obsessed with RC cars, and always wanted a teleop controller that gives us some haptic feedback of what the car is going through. This project implements that through connecting the IMU in the car with an impedance-controlled manipulator that teleops the car.
  • H.J. Terry Suh, Peter Renn, Qifan Wang, Neha Sunil,
    Manipulator-based Haptic Controller for Car-like Vehicles [pdf]
    Self-hosted, 2018.
Fast Path Planner for Car-like Nonholonomic Vehicles
Path-planning for non-holonomic vehicles can be difficult due to kinematic and dynamic constraints. We present a very fast PRM-based method to generate initial guesses for these paths using rejection sampling on steering angles.
  • H.J. Terry Suh, James Deacon, Qifan Wang,
    A Fast PRM Planner for Car-like Vehicles [pdf]
    Self-hosted, 2018.
Visual-Inertial SLAM with high-dimensional features
The goal of this project is to do SLAM with a monocular camera and an IMU. Many visual feature-tracking algorithms rely on tracking salient points across different frames, and performing sensor-fusion. We try to incorporate higher-dimensional features such as image patches and its moment dynamics to achieve better performance in structured environments.
  • H.J. Terry Suh,
    Inertial-aided Normal Estimation for Planar Feature Mapping: Unique Solutions to Homography Decomposition [pdf]
    Self-hosted, 2018.
Design of an Anthropomorphic Arm with Linear Actuators
 
What would it take to build a manipulator that closely resembles the human arm, with closed-chain linkages with linear actuators mimicking the human muscle? This work designs a human-inspired arm, and investigates kinematic properties of the arm.
  • H.J. Terry  Suh, James Deacon, Hana Keller,
    Design of an Anthropomorphic Closed-loop Chain Manipulator [pdf]
    Self-hosted, 2017.
 
Improving Off-the-shelf Localization Sensors with High-performance IMU
Recently, we have seen highly advanced off-the shelf visual odometry sensors such as Intel T265, or the Snapdragon Flight. But if we had a much better performing military-grade IMUs, could we improve the performance of these sensors using sensor fusion?
  • H.J. Terry Suh,
    Hierarchical Extended Kalman Filter with Frequent Propagation for UAV Localization [pdf]
    Self-hosted, 2017.
Humanoid-Oriented Movement Writing
 
In order to script or demonstrate full-body movements for humanoids, we need a common representation that can solve the embodiment correspondence problem. This work presents Humanoid-Oriented Movement Writing (HOM), where articulated human pose sequences are learned from videos.
  • A. Stoica, H.J.T. Suh, S.M.Hewitt, S. Bechtle, A. Gruebler, Y. Iwashita,
    Towards a Humanoid-Oriented Movement Writing [pdf]
    2017 IEEE International Conference on Systems, Man, and Cybernetics (SMC)
Intelligent Bear Robot that Learns Emotions
 
Can we design an interactive robot that predicts human actions from emotional cues and acts accordingly? We’ve designed a smart bear robot that uses a SVM to classify visual and haptic data into different emotions, and programmed it to react accordingly!
  • H.J.T. Suh, W.S. Lee, J.H. Seo,
    Design of an Intelligent Bear Robot with Emotional Learning Capabilities for Enhanced Human-robot Interaction [pdf]
    Self-hosted, 2014.

My Robots

A few people have the pleasure of designing and entire robot from scratch. It takes long hours and incredible effort to build one, but the joy of seeing it move and perform is quite incomparable to anything else. When you design and make a robot, you create some form of life; and no matter how primitive it is, you are its creator. I always imagine its blood pumping through the red and black power wires, and its brain running basic thoughts in a metal computer.

The robots I made and designed have a very special place in my heart, for I am like their father.

The Drivocopter
 

The Drivocopter was a hardware platform I built for my undergraduate thesis, funded by JPL / Caltech team’s entry in DARPA’s Subterranean challenge. It is a robot that is designed to have both driving and flying capabilities. It’s the biggest hardware project that took an year for me and the Caltech ME/EE/CS75 team to design, build, iterate, and test under guidance from Prof. Joel Burdick.

It survived many interesting crashes. I accidentally drove it into a pond one time, and a bad landing maneuver once broke all of its leg.

Here are some cool videos of the Drivocopter! [video1][video2]

Saber II: Offroad Autonomous Driving Platform
 

Saber was a off-road reconnaissance robot that a team of researchers at JPL were building with Dr. Adrian Stoica as the lead. Unfortunately, we didn’t get to finish the project due to couple hiccups right before the deadline.

I took my favorite RC car and rebooted the concept to make Saber II (though it’s missing many expensive sensors as this was a self-funded project). It uses the Traxxas X-Maxx chassis, and has an NVidia TX-2 as the central computer.

HARP: Home Autonomous Robotic Platform
 

HARP was my experiment with building a mobile platform for home robots. I never built anything of this scale (full-human size) before, so there were many interesting design (and cost!) challenges there.

Although I was never able to finish HARP, I have a video of its brief operation. [video]

The Ring of Fire: Drogon, Rhaegal, and Viserion
 

These skid-steer tank robots were built as a part Caltech’s ME72, a mechanical engineering capstone course taught by Dr. Michael Mello. Our team name was “Ring of Fire”, with members Cormac O’Neill (now a fellow grad student at MIT!), James Deacon, and Hana Keller. I was watching Game of Thrones at the time, which explains the robot names (there were three identical ones).

This is when I started to seriously delve into mechanical design, and remember spending many all-nighters trying to fix tons of failure modes. I’ll never forget the night I was pulling off an all nighter with Cormac to work on one of these robots, and one of the on-board regulators suddenly blew up in a puff of unhealthy-smelling smoke!

Ultrasound Tracking Fanboat
 
This boat was a small electronics project I built for Caltech’s ME/EE 7: Introduction to Mechatronics course. It uses an analog PID circuit to track an ultrasound signal. One might say it’s the first robot I ever made. 

Music

I am an amateur singer-songwriter. I write, record, and mix songs as a hobby, and play a lot of instruments (guitar, piano, bass, drums, harmonica, flute, etc.)
If you’re interested, check out my Soundcloud Page! (Though the last serious recording I made was in 2016)