Koopman-inspired model
Visuo-dexterity through behavioral dynamics
A Koopman-inspired visuo-dexterity model rolls out structured visual-action dynamics for temporally coherent manipulation.
Robotics PhD Student, Georgia Tech
Yunhai Han
I study structured robot learning for manipulation in complex, unstructured environments, with a focus on dexterity learning from human video, robot behavior models, and tactile manipulation.
I am advised by Prof. Harish Ravichandar at Georgia Tech. I have also been fortunate to work with Prof. Sehoon Ha, Prof. Zsolt Kira, and Prof. Danfei Xu. I received my M.S. and B.S. in Mechanical Engineering from UC San Diego and Yanshan University.
My interest in robotics started with Code Geass, and I wanted to build a mecha in high school. Now, the robots I work on are smaller than the mecha I imagined, but the motivation has stayed with me.
Research Focus
I build robot-learning systems that turn human videos, sparse demonstrations, and real-world explorative interactions into deployable manipulation skills.
Learning from human videos
Robot learning from motion and interaction priors extracted from human videos.
Visuo-dexterity learning
Learning temporally coherent visuo-dexterity skills with behavioral dynamics.
Tactile adaptation
Using touch to adapt visual policies when contact details matter most.
By integrating these directions, I aim to build open-world robot-learning agents that autonomously acquire robust dexterous skills from watching human users at deployment time and generalize across diverse operating environments, objects, tasks, and robot platforms.
Recent news
A short log of current research milestones, talks, internships, and fellowship updates.
- Research intern at NVIDIA Seattle Robotics Lab, studying humanoid robot learning from offline human manipulation videos.
- Invited talk at the Georgia Tech RoboGrads Student Seminar (poster).
- Passed the PhD proposal and advanced to PhD candidacy, with committee members Zsolt Kira, Danfei Xu, Yunzhu Li, and Dinesh Jayaraman.
- Selected as a Qualcomm Innovation Fellowship Finalist with Kelin Yu.
Robots operating in the real world
Real-world clips of Koopman-inspired visuo-dexterity, tactile policies, and autonomous robot learing from human video.
Tactile policy
Contact-aware tactile adaptation
A tactile residual policy refines a visual base policy through real-world RL, using additional tactile feedback for contact-rich tasks.
Autonomous learning from human video
One autonomous pipeline converts human manipulation videos (left) into deployable robot policies (right) across daily manipulation tasks.
Selected publications
A curated set of projects that best represent my current research direction in dexterous robot learning and tactile sensing.
OmniTacTune: Policy-Agnostic Real-World RL for Tactile Residual Adaptation of Visual Policies
Adapts tactile feedback to pretrained visual policies through real-world residual RL for contact-rich manipulation.
MimicTouch: Leveraging Multi-modal Human Tactile Demonstrations for Contact-rich Manipulation
Transfers tactile-guided human manipulation strategies to robot grippers with imitation and online residual reinforcement learning.
On the Utility of Koopman Operator Theory in Learning Dexterous Manipulation Skills
Shows that Koopman operator-based imitation learning can train dexterous manipulation skills efficiently, quickly, and with interpretable structure.
All publications
Search and filter the full publication list by title, topic, venue, author, or year.
2026
Video2Sim2Real: Full-Stack Autonomous Dexterous Acquisition from a Single Human Video
Under review
Introduces a full-stack framework that reconstructs simulator-ready digital twins from a single human video, extracts priors for robot and object motion, refines behavior in simulation, and transfers dexterous skills to real robots.
OmniTacTune: Policy-Agnostic Real-World RL for Tactile Residual Adaptation of Visual Policies
Under review
Adapts tactile feedback to pretrained visual policies through a policy-agnostic residual RL pipeline, improving real-world contact-rich manipulation across tasks, policies, and tactile representations.
Going with the Flow: Koopman Behavioral Models as Implicit Planners for Visuo-Motor Dexterity
Under review
Frames dexterous skills as coupled dynamical systems over visual and proprioceptive flows, enabling temporally coherent visuo-motor behavior through a structured Koopman representation.
2025
ImMimic: Cross-Domain Imitation from Human Videos via Mapping and Interpolation
Accepted by CoRL 2025 (Oral); Spotlight at RSS Dex Workshop
Bridges human-to-robot imitation by combining retargeted human hand trajectories, robot demonstrations, and MixUp interpolation for embodiment-agnostic co-training.
On the Surprising Effectiveness of Spectral Clipping in Learning Stable Linear and Latent-Linear Dynamical Systems
Under review
Studies a simple post-hoc spectral clipping procedure that can produce accurate, verifiably stable, and computationally efficient learned linear and Koopman dynamical systems.
AsymDex: Leveraging Asymmetry and Relative Motion in Learning Bimanual Dexterity
Spotlight at RSS Whole-body Control and Bimanual Manipulation Workshop 2025; Poster at CoRL Workshop 2024
Learns asymmetric bimanual dexterity by assigning complementary hand roles and operating over relative observation and action spaces for coordinated manipulation.
2024
KOROL: Learning Visualizable Object Feature with Koopman Operator Rollout for Manipulation
Accepted by CoRL 2024
Learns visually interpretable object features and advances them with a Koopman operator, improving manipulation from visual observations without runtime ground-truth object states.
MimicTouch: Leveraging Multi-modal Human Tactile Demonstrations for Contact-rich Manipulation
Accepted by CoRL 2024; Best Paper at NeurIPS Touch Processing Workshop
Collects multimodal human tactile demonstrations and uses imitation plus online residual RL to transfer tactile-guided strategies to robot grippers.
2023
On the Utility of Koopman Operator Theory in Learning Dexterous Manipulation Skills
Accepted by CoRL 2023 (Oral)
Shows that Koopman operator-based imitation learning can train dexterous manipulation skills efficiently, quickly, and with interpretable structure.
2022
Leveraging Heterogeneous Capabilities in Multi-Agent Systems for Environmental Conflict Resolution
Accepted by SSRR 2022
Introduces a temporal-logic-based controller synthesis framework for heterogeneous agents that resolve runtime environmental conflicts while preserving safety and task completion.
2021
Autonomous Vehicles for Micro-mobility
Accepted by Autonomous Intelligent Systems, Springer 2021
Summarizes the design, deployment lessons, and system improvements from autonomous micro-transit and mail-delivery vehicle work on the UC San Diego campus.
2020
Students and collaborators
Students I have mentored and external collaborators I have worked with across robotics, control, and robot learning.
Mentored students
- Kelin Yu CS M.S., 2022-2024; now PhD at UMD CS
- Zhenyang Chen Robotics M.S., 2023-2024; now PhD at GT Robotics
- Hanyao Guo Robotics M.S., 2024-2025; now SDE at Amazon
- Pratik Shah CS M.S., 2024-2025; now MLE at Amazon Annapurna Labs
- Yuqian Zheng CSE M.S., 2024-2025
- Yangcen Liu Robotics M.S. in Prof. Danfei Xu's lab, 2024-2025; now CS PhD at UChicago
- Woochul Shin CS M.S. in Prof. Danfei Xu's lab, 2024-2025; now CS PhD at UMD
- Linhao Bai CS M.S., 2024-present
- Ziyu Xiao ECE M.S., 2025-present
- Jianuo Qiu Robotics M.S., 2026-present
- Jiaqi Fu Robotics M.S., 2026-present
- Zihang Zeng Robotics M.S., 2026-present
External collaborators
- Jeonghwan Kim GT CS PhD, Prof. Sehoon Ha's lab
- Hongyi Chen CMU Robotics PhD, Prof. Jeffrey Ichnowski's lab
- Manisha Natarajan GT Robotics PhD, Prof. Matthew Gombolay's lab
- Muhammad Zubair Irshad GT Robotics PhD, Prof. Zsolt Kira's lab
- Kenneth Shaw CMU Robotics PhD, Prof. Deepak Pathak's lab
- Sizhe Wei GT Robotics PhD, Prof. Lu Gan's lab
Group projects
Hands-on robotics projects that shaped how I think about perception, control, and deployed systems.
RoboMaster
RoboMaster is a DJI-founded platform for robotic competitions and academic exchange. I was the vision group leader of YSU Eagle, where my group worked on object tracking, range estimation, serial-port communication, and gimbal stability control for mobile robots.
Service, teaching, and interests
Academic service, teaching experience, and a few personal interests outside the lab.
Professional service
- GT RoboGrads Research Vice President
- Reviewer ICRA, IROS, RSS, CoRL, RA-L, IEEE T-RL, NeurIPS, ICLR, AIM, SSRR
- Session Chair ACC
Teaching
- MAE145, Robotic Planning & Estimation, UCSD Teaching Assistant, Winter 2021
- MAE146, Introduction to ML Algorithms, UCSD Teaching Assistant, Spring 2021
Working experience
- Research Assistant, Georgia Institute of Technology Summer 2021 to Spring 2022
Misc
- Running Finished a half-marathon within 2 hours
- Soccer and badminton
- Photography
- Gaming League of Legends
- Anime, manga, and movies