Younghyo Park

Younghyo Park is an EECS Ph.D. student at MIT CSAIL, advised by Pulkit Agrawal, and a member of the Improbable AI Lab. His work examines practical choices that shape robot learning: how training environments are designed, how demonstrations are collected and how low-level controllers interact with learned policies.[2][4][5][10][8]

Park and Antonia Bronars contributed equally to Tune to Learn: How Controller Gains Shape Robot Policy Learning, coauthored with Agrawal and submitted in April 2026; their author order was determined by coin flip. The study asks how to choose position-controller gains when a learned policy supplies the commands. Its central argument is that effective stiffness depends on both controller dynamics and the policy’s learned reactions, so gains should be selected for how well they support learning.[8]

Across the tasks and robot embodiments they studied, the authors report that behavior cloning benefits from compliant, overdamped gain regimes; reinforcement learning can succeed across gain regimes with compatible hyperparameter tuning; and stiff, overdamped regimes harm transfer from simulation to real robots. These results make gain selection dependent on the learning method. Whether the trends extend to cross-embodiment learning or whole-body control remains an open question.[8][9]

In his research vision, updated in April 2026, Park writes, “I believe robot learning is fundamentally different from language or vision.” He describes robots as coupled systems of hardware, controllers, sensors, data collection interfaces and learning algorithms, where changing one component changes how the others behave. His stated aim is to use open-source infrastructure and controlled empirical studies to make robot training more principled and reproducible.[2]

Earlier MIT projects addressed two other parts of this learning pipeline: the human effort required to prepare reinforcement-learning environments and the difficulty of collecting robot demonstrations at scale.[5][10]

At ICML 2024, Park coauthored Position: Automatic Environment Shaping is the Next Frontier in RL with Gabriel B. Margolis and Agrawal, sharing equal-contribution credit with Margolis. They argued that a central bottleneck in robotic reinforcement learning is the labor of designing observations, actions, rewards and simulation dynamics to make a task learnable. They called for research that automates this environment shaping.[5][2][6]

Their project explanation makes the scope concrete. Action-space shaping determines how a policy’s output becomes a motor command. Reset strategies determine physically feasible starting conditions: randomly perturbing object positions can produce overlapping geometry, while dropping objects into a bin lets simulated physics generate a cluttered arrangement. These examples illustrate the task-specific engineering the authors wanted to automate.[6]

Park also coauthored DexHub and DART: Towards Internet Scale Robot Data Collection with Jagdeep Singh Bhatia, Lars Ankile and Agrawal, submitted in November 2024. DART combines cloud-based simulation with augmented-reality teleoperation. Visual rendering runs locally on the AR device, hand tracking supplies robot commands, and pinching and releasing fingers controls parallel-jaw grippers. Instant randomized resets reduce the need to rearrange a physical scene between demonstrations; collected data is automatically logged to DexHub.[10][11]

In the team’s reported study, twenty robotics novices collected demonstrations in seven-minute sessions using DART and the Rainbow RB-Y1 and ALOHA default “kinematic double” teleoperation interfaces. Participants collected 2.5 times as many demonstrations on average with DART. The authors also report lower physical fatigue and successful transfer of policies trained on DART data to real robots, including robustness to unseen visual disturbances.[11][10]

Park’s practical work includes VisionProTeleop, credited with Agrawal, which streams Apple Vision Pro hand and head tracking to robotics systems and returns video, audio or simulation feedback. In Winter IAP 2025, he taught the Modern-Robot Learning hands-on tutorial with Haoshu Fang and Agrawal, covering data collection, policy training and evaluation in simulated environments.[2]

Before joining MIT in 2023, Park worked on manipulation learning at NAVER LABS.[2][7]

Park earned a bachelor’s degree in mechanical engineering, summa cum laude, at Seoul National University. His CV lists a full-time robotics-engineer role at NAVER LABS from March 2022 to July 2023, researching and developing unsupervised skill-discovery algorithms for manipulation. His personal biography describes developing machine-learning and reinforcement-learning algorithms for robot arms such as AMBIDEX to perform daily tasks. He left NAVER LABS in 2023 to begin doctoral work with Agrawal at MIT CSAIL.[2][7]