Minwoo Cho

I am an M.S. student in Computer Science at KAIST, specializing in Reinforcement Learning (RL) and Robotics. Under the supervision of Prof. Daehyung Park in the RIRO (Robust Intelligence & RObotics) Lab, I am currently working on learning constraints—such as safety conditions and task specifications—from demonstrations. By incorporating these constraints into reinforcement learning, I aim to develop policies that are both generalizable and safe in real-world applications.

I received my B.S. from KAIST in 2023, where I majored in Mechanical Engineering and minored in Mathematical Sciences.

Email  /  CV  /  Github

Research Interest

My research focuses on applying RL to real-world robotics, where several fundamental challenges such as uncertainty or sample efficiency still hinder deployment on physical robots. To address these challenges, my current research interests include:

• RL with constraints: Unlike rewards, constraints can effectively restrict unintended or unsafe behavior, which is crucial for real-world deployment. However, these constraints are often difficult to explicitly define and ensuring the learning of zero-violation policies in complex domains remains challenging. How can we effectively learn and leverage such constraints?


• Offline-to-Online RL (O2O): Since collecting real-world data is expensive, O2O RL methods offer a promising approach to improve sample efficiency and bridge the sim-to-real gap. Further reducing the number of online interactions and applying these methods in safety-critical settings is essential for real-world deployment.


• Learning from video: Learning task knowledge from videos can eliminate the need for handcrafted rewards in RL. Just as humans learn by observing, I'm interested in enabling robots to acquire high-quality skills from a diverse range of videos without requiring task-specific designs.

Publications

In this work, we learn free-form temporal logic constraints from demonstrations using a tree-based genetic algorithm and constraint-regularized RL with logic cost redistribution.

In this work, we perform semantic navigation using a scene-graph grounding network that predicts the object from a natural language input, on RBQ-3 quadruped robot.

Projects

This project aims to learn RL policies for manipulating everyday objects in diverse rigid and deformable scenarios—such as pick-and-place and entangling/disentangling—directly from visual input. We build on SERL as the code base.

This project aims to infer common constraints from multi-task demonstrations by decomposing the task-agnostic part of the reward inferred from Meta-IRL.

This project aims to generate a navigation plan based on the traversability prediction network. Traversabilities are estimated from the voxels of the terrain and trained with the locomotion performance of RL policies in various terrains.

This project aims to design and create a self-driving hovercraft to navigate to given goals using a 2D LiDAR and BLDC motors. We designed the hovercraft with two thrust propellers, and controlled the machine using PD control.

This project aims to creatively design a Goldberg machine in a Gazebo simulator. Enjoy the video!

Template from Jon Barron.