Reinforcement learning for robot grasping
Master thesis- Axis:
- Transition sociétale et technologique des entreprises
- Site :
- Toulouse
- Type:
- Master thesis
- Supervising person(s):
- Philippe Juhel
We want the robot to automatically grasp objects from a bunch of random positioned objects, with no previous knowledge about them.
To do that, the robot will perform some trials and learn the success grasping configurations.
We want to study how we can improve the training speed (i.e. do less trials with the same success grasping rate).
We use a Universal Robot UR3 and a home made vacuum gripper.
The technologies that you will use are :
- Linux Ubuntu 20.04 operating system
- ROS (Robotic Operating System) : the very well known and used robotics middleware for all kinds of robotics projects.
- deep learning (DL) and reinforcement learning (RL) : techniques to train a system from many examples.
- pytorch and lightning : AI python libraries to program DL and RL softwares
OpenCV : python computer vision library
Two master thesis have been done before and you will continue these works.
The main goal is to search for new ideas to improve the training speed and/or the success/fail grasping rate.
Another subject is to take into account a multiflash camera to study if it can improve the results.
- required skills:
-
You must have some programming skills (preferably in Python).
Skills on ROS or AI (Deep learning and/or Reinforcement Learning) would be preferable but not mandatory.
You MUST be curious and not afraid to learn new technologies.