I want a world where robots can perform every dexterous manipulation task a five-year-old can do with their two hands — and where building such a system is genuinely easy, not a multi-year research effort. Some argue we can get there by following the same playbook that drove rapid progress in vision and language. I'm not so sure.

I believe robot learning is fundamentally different from language or vision. A robot is inherently interdisciplinary — a tightly coupled system of hardware, controllers, sensor configurations, data collection interfaces, and learning algorithms — and changing any one of them rewires how the others behave. This makes robotics much closer to biology than to modern ML: the real scientific question isn't just which model is best, but how these many interacting pieces shape what a robot can ultimately learn. And we're still in the early days of building that kind of science.

My research aims to build this missing science. I develop open-source infrastructure that the community uses to collect and process robot data at scale, and I use that infrastructure to run controlled empirical studies that reveal how under-examined design choices — from low-level controller gains to data collection interfaces — fundamentally determine learning outcomes. The goal is a future where training a robot to perform a new task is less art and prayer, more engineering: principled, scalable, and reproducible.

This course provides a practical introduction to training robots using data-driven methods. Key topics include data collection methods for robotics, policy training methods, and using simulated environments for robot learning. Throughout the course, students will have hands-on experience to collect robot data, train policies, and evaluate its performance.

• [2025 Nov 07] Prof. Sunghoon Ahn's lab (IDIM), Seoul National University
• [2024 Aug 07] Prof. Kimin Lee's lab, KAIST

I'm passionate about building open-source tools that empower the robotics community and streamline the robot development experience.

A complete ecosystem for using Apple Vision Pro in robotics research — stream hand/head tracking from Vision Pro, send video/audio/simulation back for real-world teleoperation, simulation teleoperation, and egocentric dataset recording.

Run your own custom 1 kHz torque controller on a Franka, right from your Mac (or Linux): write it in Python, and aiofranka runs it in a C++ loop that never waits for Python. Plus helpful commands so you never have to touch the Desk GUI: unlock and lock the robot, run self-tests, hand-guide it, identify payloads, and calibrate cameras.

The aiofranka counterpart for Flexiv robots (Rizon 4 and 4S). An asyncio-based Python library that pairs flexivrdk with a small pybind11 shim for real-time joint-torque control, offering joint impedance, operational-space, and direct torque control, smooth Ruckig moves, and a MuJoCo backend for payload-aware dynamics and simulation.

Generate 3D-printable cubes with ArUco or AprilTag fiducial markers on all six faces, then detect their full 6-DoF pose from a single camera image. A two-part pipeline: a generator that produces multi-color 3MF files ready for dual-color 3D printing, and a detector that estimates rotation and translation given camera intrinsics.

* denotes equal contribution.

Robot arms are often far heavier than the everyday objects they manipulate. Through controlled experiments that vary robot and object masses, we show that this inertia mismatch carries a hidden cost for behavior cloning, and that choosing workspace placements and demonstrations that favor lower task-space inertia improves imitation-learned policies without replacing the robot.

A blind insertion policy that senses contact only through its own joint encoders, with no vision, force, or tactile input, is trained entirely in simulation and transfers zero-shot to a real robot, inserting pegs of diverse shapes and sizes with millimeter-scale clearance. Blind policies are slower than matched visual ones, but generalize far better to unseen parts.

* equal contribution

We show that controller gains shape the inductive bias of different policy learning paradigms, and identify gain regimes that maximize learnability for behavior cloning, reinforcement learning, and sim-to-real transfer.

* equal contribution, order determined by coin flip

DART is a teleoperation platform that leverages cloud-based simulation and augmented reality (AR) to revolutionize robotic data collection. It enables higher data collection throughput with reduced physical fatigue and facilitates robust policy transfer to real-world scenarios. All datasets are stored in the DexHub cloud database, providing an ever-growing resource for robot learning.

Most robotics practitioners spend most time shaping the environments (e.g. rewards, observation/action spaces, low-level controllers, simulation dynamics) than to tune RL algorithms to obtain a desirable controller. We posit that the community should focus more on (a) automating environment shaping procedures and/or (b) developing stronger RL algorithms that can tackle unshaped environments.

* equal contribution

An algorithm that can discover diverse and useful set of skills from scratch that is inherently safe to be composed for unseen downstream tasks. Considering safety during skill discovery phase is a must when solving safety-critical downstream tasks.

* equal contribution

Drawing robot ARTO-1 performs complex drawings in real-world by learning low-level stroke drawing skills, requiring delicate force control, from human demonstrations. This approach eases the planning required to actually perform an artistic drawing.

* equal contribution

We detect collisions for robot manipulators using unsupervised anomaly detection methods. Compared to supervised approach, this approach does not require collisions datasets and even detect unseen collision types.

Check out Jon Barron's repository for the template of this website.