Dong-Ki Kim

I am a PhD student at MIT-LIDS advised by Professor Jonathan P. How. My research focuses on the field of reinforcement learning and robotics. Specifically, I am interested in multi-agent reinforcement learning for learning to coordinate with other simultaneously learning robots, meta-learning for enabling a robot to adapt fast to unseen situations, and hierarchical learning for solving the delayed credit assignments.

Previously, I received my B.S. in Electrical and Computer Engineering (summa cum laude) at Cornell. Before MIT, I have been advised by such wonderful advisors: Professor Sebastian Scherer at CMU-RI, Professor Matthew R. Walter at TTIC, and Professor Tsuhan Chen at Cornell.

Email  /  CV  /  Google Scholar  /  GitHub



Reinforcement Learning

PontTuset Influencing Long-Term Behavior in Multiagent Reinforcement Learning
Dong-Ki Kim, Matthew Riemer, Miao Liu, Jakob N. Foerster, Michael Everett, Chuangchuang Sun, Gerald Tesauro, Jonathan P. How
ICLR-22 Workshop (Spotlight)

In this paper, we propose a new multiagent optimizaton for considering the limiting policies of other agents as the time approaches infinity.

PontTuset Context-Specific Representation Abstraction for Deep Option Learning
Marwa Abdulhai, Dong-Ki Kim, Matthew Riemer, Miao Liu, Gerald Tesauro, Jonathan P. How
Paper / Code / Video

We introduce Context-Specific Representation Abstraction for Deep Option Learning (CRADOL), a new framework that considers both temporal abstraction and context-specific representation abstraction to effectively reduce the size of the search over policy space.

PontTuset A Policy Gradient Algorithm for Learning to Learn in Multiagent Reinforcement Learning
Dong-Ki Kim, Miao Liu, Matthew Riemer, Chuangchuang Sun, Marwa Abdulhai, Golnaz Habibi, Sebastian Lopez-Cot, Gerald Tesauro, Jonathan P. How
ICML-21, AAAI-20 Symposium
Paper / Code / Video

We develop a novel meta-multiagent policy gradient theorem that directly accommodates for the non-stationary policy dynamics inherent to multiagent settings. Our meta-agent directly considers both an agent’s own non-stationary policy dynamics and the non-stationary policy dynamics of other agents to adapt fast.

PontTuset FISAR: Forward Invariant Safe Reinforcement Learning with a Deep Neural Network-Based Optimizer
Chuangchuang Sun, Dong-Ki Kim, Jonathan P. How
ICRA-21, ICML-20 Workshop

We propose to learn a neural network-based meta-optimizer to optimize an objective while satisfying constraints, where the constraint satisfaction is achieved via projection onto a polytope formulated by multiple linear inequality constraints.

PontTuset Learning Hierarchical Teaching in Cooperative Multiagent Reinforcement Learning
Dong-Ki Kim, Miao Liu, Shayegan Omidshafiei, Sebastian Lopez-Cot, Matthew Riemer, Golnaz Habibi, Gerald Tesauro, Sami Mourad, Murray Campbell, Jonathan P. How
AAMAS-20, AAAI-19 Workshop
Paper / WIRED News

We introduce a new learning to teach framework, called Hierarchical MultiagentTeaching (HMAT). Our framework solves difficulties faced by previous learning to teach works when operating in domains with long horizons, large state spaces, and continuous actions.

PontTuset Policy Distillation and Value Matching in Multiagent Reinforcement Learning
Samir Wadhwania, Dong-Ki Kim, Shayegan Omidshafiei, Jonathan P. How
Paper / Video

We introduce a multi-agent algorithm that combines knowledge from agents through distillation and value-matching (DVM). DVM outperforms policy distillation alone and allows faster learning in dynamic tasks.

PontTuset Learning to Teach in Cooperative Multiagent Reinforcement Learning
Shayegan Omidshafiei, Dong-Ki Kim, Miao Liu, Gerald Tesauro, Matthew Riemer, Christopher Amato, Murray Campbell, Jonathan P. How
AAAI-19 (Outstanding Student Paper Honorable Mention), ICML-18 Workshop
Paper / MIT News

This paper presents Learning to Coordinate and Teach Reinforcement (LeCTR), the first general framework for intelligent agents to learn to teach in a cooperative MARL.

PontTuset Crossmodal Attentive Skill Learner
Shayegan Omidshafiei*, Dong-Ki Kim*, Jason Pazis, Jonathan P. How
JAAMAS-20, AAMAS-18, NeurIPS-17 Symposium
JAAMAS-20 Paper / AAMAS-18 Paper / Video / Code

This paper introduces the Crossmodal Attentive Skill Learner (CASL), integrated with the recently-introduced Asynchronous Advantage Option-Critic architecture to enable hierarchical reinforcement learning across multiple sensory inputs.

Computer Vison and Robotics

PontTuset Demonstration-Efficient Guided Policy Search via Imitation of Robust Tube MPC
Andrea Tagliabue, Dong-Ki Kim, Michael Everett, Jonathan P. How
Paper / Video

By leveraging properties from a robust tube model predictive controller, we introduce a data augmentation method that enables high demonstration-efficiency regardless of perturbations.

PontTuset Decentralized Multi-UAV Package Delivery
Dong-Ki Kim, Jesús Tordesillas Torres, Hao Shen, Samir Wadhwania, Sebastian Lopez-Cot, Alan Osmundson, Hector Castillo, Jesko (Diego) Mueller, Brett Lopez, Shayegan Omidshafiei, Jonathan P. How
Boeing Annual Visit-18

For the Boeing annual visit at MIT, we prepared a demo, a package delivery scenarios with multiple drones. My contributions include the RVO-based collision avoidance, on-board perception system for delivery type classification, and projection system for visualization.

PontTuset Satellite Image-based Localization via Learned Embeddings
Dong-Ki Kim, Matthew R. Walter
Paper / Video / NVIDIA News

We propose a vision-based method that localizes a ground vehicle using publicly available satellite imagery as the only prior knowledge of the environment.

PontTuset Autonomous Off-Road Vehicle
CMU-RI and Yamaha Team-17
Video1 / Video2 / Video3

We developed an autonomous off-road vehicle (video1). My contributions to the project include a ROS-based system that estimates terrain roughness from 3D point cloud in real-time (video2, video3).

PontTuset Season-Invariant Semantic Segmentation with A Deep Multimodal Network
Dong-Ki Kim, Daniel Maturana, Masashi Uenoyama, Sebastian Scherer

We propose a novel multimodal architecture consisting of two streams, image (2D) and LiDAR (3D). By combining the two streams, we achieve a robust season-invariant semantic segmentation.

PontTuset You Are Here: Mimicking the Human Thinking Process in Reading Floor-Plans
Hang Chu Dong-Ki Kim, Tsuhan Chen
Paper / Video

We address the problem of locating a user in a floor-plan, by using a camera and a floor-plan.

PontTuset Deep Neural Network for Real-Time Autonomous Indoor Navigation
Dong-Ki Kim, Tsuhan Chen
Technical Report-15
Paper / Video

We propose a vision-based deep learning system in which a drone autonomously navigates indoors and finds a specific target.

Source code credit to Dr. Barron