Projects @ UC Berkeley:
Deep Reinforcement Learning

Deep reinforcement learning with the Advantage Actor-Critic (A2C) model. In addition to standard A2C, proximal policy optimization (PPO) is also implemented. As discussed in the original paper, PPO provides better stability by preventing the policy from changing too much during each update. The models are evaluated on multiple tasks from OpenAI gym. Code available soon!

Neural Artistic Style

Here is a neural network that creates stunning artistic images! This project replicates the result in "A Neural Algorithm of Artistic Style", Gatys et al. 2015. A deep neural network is trained using TensorFlow to create new artistic images by recombining styles and contents from different inputs. Code available on GitHub.

The neural network used in this project is based on a VGG16 model pre-trained on ImageNet. The model is available here.

Neural Poet 

Use a Recurrent Neural Network (RNN) to generate classic Chinese poems! This project uses a Sequence-to-Sequence model inspired by this TensorFlow tutorial.

Currently, the model has learned from over 200,000 lines of original poems. Check out this interactive web app:

http://neural-poet.herokuapp.com/

Code available soon. Enjoy!

Agile Robot Navigation with Floor Plan

One thing that prevents a robot from navigating in an indoor space is that it lacks an accurate map. While we can certainly use SLAM to generate such a map, this is extremely costly as the robot needs to scan the space every time it arrives in a new environment. In this project, I combined particle filtering with AR tag recognition so that the robot can navigate with an inaccurate floor plan.

Robotic Platform for Air-Change Effectiveness Assessment 

Standard approach for measuring air-change effectiveness requires multiple sensor stations to be installed. The calibration and deployment of these sensors become a daunting task, especially when the space is big. In this project, we propose and evaluate an alternative profiling method of indoor ventilation efficiency based on autonomous robotic platform.  [Paper] [Poster]

 
CONTACT ME

Yulun Tian

Graduate student at MIT AeroAstro

Email:

yulun [at] mit [dot] edu

  • LinkedIn - Black Circle
  • github-10-xxl
  • Facebook - Black Circle
  • YouTube - Black Circle

© 2020 By Yulun Tian.