Yuki Noguchi

I'm a researcher at Toyota's Frontier Research Center involved in robot learning R&D. Before that, I was a one year trainee at Toyota Research Institute and a master's student at Matsuo Lab, part of the TRAIL subgroup, at the University of Tokyo.

I'm interested in general-purpose robots for both everyday and industrial use. I believe that simple but scalable data-driven methods present the most promising path towards this.

In high school, I developed games as a hobby. During my bachelor's, I worked with computer vision and reinforcement learning. Since then my interests have shifted and grown to include robotics.

I was born and raised in LA. I've lived in California for 19 years and Japan for 10.

GitHub | Resume | LinkedIn

Personal email: in resume

Projects

Some of the things I've worked on, some available on GitHub. Roughly in order of recency.

AlohaMini (2025)

Another low-cost mobile manipulator built by following an open-sourced project here. So far I've implemented a teleop system using a Meta Quest. More details on X here.

Low-cost mobile manipulator (2024)

A Koch arm mounted on a mecanum wheel base with iPhone-based localization. It can be controlled with an end-to-end policy trained on demonstrations collected with an UMI-like device. More details on X here: 1 2 3.

Imitation learning demo (2021)

A little demo video of a CNN-LSTM policy trained to place a bowl on a plate using imitation learning. For the training data I collected about 100 demos, which took around an hour, using a teleop interface that I developed from an unused Vive lying around in the lab. The positions of the plate and bowl were shifted randomly during data collection, allowing the policy to deal with some variation as shown in the video. This was when I was first personally awed by imitation learning for real-world robotics. Two years later, the training data was merged into Open-X Embodiment.

Tool as Embodiment for Recursive Manipulation (2021)

A research project toward unifying hand-object and tool-object manipulation, part of my master's thesis. Honestly, the biggest takeaway was that it's pretty hard to model complex physical interactions in simulation without hacks. Arxiv

Sim-to-real image segmentation and RL grasping

A PyBullet simulation along with code for training an image segmentation model and an RL-based grasp policy. This was developed for the competition described below.

World Robot Summit 2020 and RoboCup Japan Open 2020 (2019-21)

Two robot competitions that used the HSR, Toyota's mobile manipulator robot, for household cleanup and retrieval tasks. I was in Team Weblab and mainly wrote the perception and grasping modules of the system. There was much trial and error during development, like trying various grasp planning methods, implementing a Unity-based VR teleop interface, and using sim-to-real. I also made a crude vision-based tactile sensor (a red fingertip) for grasp success checking. This was a project that spanned 2 years in which I learned about the many challenges of real-world robotics. Sample Video, Arxiv, Advanced Robotics

DQN implementation in TensorFlow.js (2018)

A browser-based DQN inspired by an old ball-dodging iPhone game.

sumica: Activity Recognition for Home Automation (2017-18)

A home automation software that features vision-based action recognition and visual programming. This project was developed with a former colleague under the MITOU program.

Claw game environment (2017)

A claw game environment in Unity that was used for reinforcement learning experiments and a small hackathon. Video

Progressive Neural Networks experiment (2016)

A Progressive Neural Networks (a continual learning method) implementation used for an agent in a custom Unity-based RL environment. Made for a hackathon.

SSD implementation in TensorFlow (2016)

Back when there weren't so many DL object detection libraries.

Hard attention implementation (2015)

An implementation of the Recurrent Attention Model.