Taerim Yoon

I am a postdoctoral researcher at ETH Zurich 🇨🇭, jointly affiliated with the Computational Robotics Lab (CRL) led by Prof. Stelian Coros and the Computer Vision and Learning Group (VLG) led by Prof. Siyu Tang. I received my Ph.D. in Artificial Intelligence from Korea University 🇰🇷, advised by Prof. Sungjoon Choi.

My research aims at lifelike robots that live and work in human-centered environments. I develop robots that learn physical skills from the movements of animals and people and act according to those around them, spanning legged locomotion, whole-body dexterity, and multimodal human-robot interaction.

Email  /  Scholar  /  Twitter  /  LinkedIn

We learn an invertible mapping that warps the configuration space into a latent space where the collision-free region becomes a union of convex polytopes. Planning over these explicit convex sets with a Graph of Convex Sets solver yields fast, feasibility-preserving paths, and the explicit boundaries can be refined at test time. Demonstrated on 2D navigation, 6-DoF reaching, and real-time 14-DoF bimanual manipulation.

A human-in-the-loop framework that teaches legged robots agile navigation behaviors from luring, gestures, and speech, the way one would train a dog. Interaction scenes are reconstructed in physics-based simulation to augment limited demonstrations, and a progressive goal cueing strategy aligns the learned behaviors with multimodal human commands.

GRIT is a two-stage framework that learns dexterous grasping from sparse taxonomy guidance: it first predicts a taxonomy-based grasp specification from the scene and task, then a policy conditioned on this sparse command generates continuous multi-finger control. This keeps the controller steerable by users while avoiding dense pose or contact targets for every object and task.

We present a modular high-DOF tendon-driven soft finger and customizable soft hand system capable of diverse dexterous manipulation tasks. By integrating an all-in-one actuation module and enabling flexible finger arrangements, our design supports versatile, task-oriented soft robotic platforms.

Our method enables terrain-aware motion retargeting and time-critical skills such as BackFlip and HopTurn from noisy inputs, including videos or physics-ignorant kinematic frames, and successfully deploys them on real robots.

APEX is a plug-and-play extension to motion tracking algorithms that integrates expert demonstrations into reinforcement learning through decaying action priors. It removes any dependence on reference data at deployment, improves sample efficiency, and reduces tuning effort, yielding robust, animal-like locomotion on legged robots.

Our method enables terrain-aware motion retargeting and time-critical skills such as BackFlip and HopTurn from noisy inputs, including videos or physics-ignorant kinematic frames, and successfully deploys them on real robots.

Our goal is to find the right level of autonomy between humans and robots. Fully autonomous control can reduce user controllability, while fully manual control can be burdensome. To address this, we propose a framework that allows the robot to suggest a diverse set of high-quality trajectory options, enabling the user to guide the robot more effectively.

We propose a kinematics-informed neural network (KINN) that enables safe, dexterous, and unified control of hybrid rigid-soft robots by combining rigid body priors with data-driven learning.

This work introduces Zero-shot Active Visual Search (ZAVIS), a system that enables a robot to search for user-specified objects using free-form text and a semantic map of landmarks. By leveraging commonsense co-occurrence and predictive uncertainty, ZAVIS improves search efficiency and outperforms prior methods in both simulated and real-world environments.

Projects

Unpublished, yet interesting projects that I have worked on.

Using reinforcement learning, we train a quadruped robot to perform parkour skills such as jumping over and between obstacles.