Baris Yazici

Learning object grasping from raw depth sensor data with Reinforcement Learning in Pybullet simulation.

This project has begun with my master’s thesis project- full report can be foud here. I developed the project even further after I handed my dissertation. It evolved to a fully open-source project that aims to help all robotics enthusiast. This projects presents a convenient way to train robotics hand for object manipulation. It includes two different scenes(table and plane) to test the trained model. It uses OpenAI-Gym to describe the environment. It has built-in support for depth, RGB-D and auto-encoder observation. There are trained models in SAC, DQN and in BDQ Reinforcement Learning algorithm.

The trainning structure supports curriculum learning, shaped reward integration for speeding up the training.

Controlling the tendons of the Roboy with PPO algorithm. Trained the agent on in a Docker container

Roboy’s joints are actuated with tendons, akin to humans. It is challenging to control the length of tendon to build to move the joint in a specified problem. The traditional control approach suffers to perform, because of the modelling difficulties, such as elasticity of the tendon or the uncertainty of the tendon length when it is wrapped around the pulley.

Therefore we developed an adaptable system with Deep Reinforcement Learning. We describe the environment with OpenAI gym and trained the model with PPO algorithm on Google Cloud Platform. I implemented the gym interface and the physics simulations helper functions.

Term project of 3D Scanning and Motion Capture course at TUM. 3D Reconstructed the world and interacted with it in Unity

Developed a real-time kinect fusion pipeline for reconstruction and augment reality application. We used Iterative Closest Point(ICP) algorithm for camera tracking and marching cubes to extract the meshes. Later we transferred the meshes to Unity and interacted the world with such as rendering balls or bottles on the reconstructed world.

Detect hand gestures with SSD transfer to Gazebo simulation via ROS

Roboy is one of the biggest student teams in Technical University of Munich. We are developing a tendon-actuated humanoid robot. One part of the project was to detect the hand with SSD model in a real-time video stream from Leap Motion camera. The other part is the classify the hand with a specific gesture-id. This gesture is then passed to the simulation of the Hand model via ROS communication.

Hand model is develop based on a previous template on AutoDesk Fusion

Web service to facilitate the communication between extreme sporters. Developed on Ruby on Rails and deployed with heroku

I started this project to familiarize myself with web development in the early stages of my university education :) The website is based on RESTFUL api. Every user can add a category(accomodation, transport, guidance) for a specific extreme sport type(Rock Climbing, Ski Touring or Wind Surf). It allows user to search functionality with elasticsearch plugin.

Fleet Management application that makes use of BMW Car-Data API. Developed with Node.js and MongoDB in the backend and react in the front-end

As part of Technical Unviversity Munich’s mobility services course in coordination with BMW, we developed a Fleet Management application using BMW CarData. We used MongoDB, Express.js and Node.js in the backend and React for the front-end.