2018 Intern Results at Preferred Networks (Part 1)

This summer, Preferred Networks accepted a record number of interns in Tokyo from all over the world. They tackled challenging tasks around artificial intelligence together with PFN mentors. We appreciate their passion, focus, and designation to the internship.
In this post, we would like to share some of their great jobs (more to come).

Learning to flip objects using guided policy search

Takeshi Itoh (Nara Institute of Science and Technology)
Control theory for general object handling is essential for applying robotics to a broader range of tasks. Although dexterous manipulation with robotic hands has been intensively studied, most of those studies focused on manipulating a single object and did not consider generalizing their control theories to various objects. In this study, we focus on flipping manipulation because it requires capabilities to cope with friction, slip, and contact, which are fundamental for various kinds of dexterous manipulation such as regrasping. We demonstrate that a 3-DoF manipulator can learn a policy to flip an object on a table using guided policy search, and that this policy can be generalized to flip other objects which have same shapes but with different colors. As a further step, we are now trying to generalize this policy to objects with different shapes.


 

Estimation of baby’s emotion and intention

Marina Y. Aoyama (University of London)
We challenged representation learning of baby faces for estimation of their emotion and intent. We trained CNN-based models on images on the Web, analyzed their clusters, and built a real-time demo system with a camera.

 

Replication Study and Extension to Transfer Learning of “Emergence of Locomotion Behaviors in Rich Environment”

Manabu Nishiura (University of Tokyo)
Recently, deep reinforcement learning (DRL) has achieved great success in challenging legged robot locomotion. Especially, Heess et. al. (https://arxiv.org/abs/1707.02286) shows that a diverse set of challenging terrains and obstacles help DRL agent to learn run, jump, crouch, and turn as required by the environment by using only simple reward function. In this internship, we built experiment environment to replicate their paper. As an extension of the research, we are trying to reduce learning time by using transfer learning among different bodies.




 

Realtime generating sounds with Wavenet on FPGA

Ryoma Sato
His theme of this internship is to accelerate wavenet with FPGA. His ultimate goal is realtime generating sounds only with wavenet, which was accomplished in the end of intern period.


(Original sound: Baby Just Born Sound, recorded by Daniel Simon, Creative Commons Attribution 3.0)
—
We are planning to expand the summer internship program next year. Though the 1st set of applications for students outside of Japan are already in the review process, we will open the next application later. Please follow us on social media for updates.