MVA Projects

Robotic Arm

This project aimed at learning to control a vision based robotic arm through imitation of an expert. Such a situation is simulated using the “FetchPickAndPlace” environment from OpenAI. To solve the task, the arm should grab the cube and take it to the red target.

An expert can be easily programmed using the 3D positions of objects. It is then interesting to learn a policy that does not have access to those positions and instead relies on vision. With enough graphical variations we could even hope that this policy could perform well in the real world.

Our work was based on an implementation of Behavioral Cloning from Robin Strudel. Our contribution was to implement and experiment with DAgger and DART which are Imitation Learning algorithms. The Code and Scientific Report are available.

Graph Based Reinforcement Learning

During this project, we designed a Reinforcement Learning environment that involved graphs and experimented with novel Deep Neural Network architectures that exploit the graph structure. The environment was basically a grid with walls and a vacuum cleaner that should move around and clean dust. For more details, check out the Code and Scientific Report.

Breast Cancer Metastases Detection

This is participation to a Data Challenge provided by Owkin. The goal was to detect weather a patient contracted brest cancer metastases or not given about 1,000 samples. For more details, check out the Code and Scientific Report.