In the project, you will work with robot models in simulated environments, e.g. [1, 2, 3, 4], and build a controller for manipulation of objects using general-purpose deep reinforcement learning algorithms [5], modern neural network architectures and classic approaches. You will experiment with adding touch sensors [6] to the hand and explore generalization of learned policies over different objects with and without tactile information.
[1] OpenAI Robotics
http://gym.openai.com/envs/#robotics
https://drive.google.com/open?id=1J2H92AstGpcFYqjmVymMSRKI5xISdB1b
[2] Robot open-Ended Autonomous Learning https://www.aicrowd.com/challenges/neurips-2019-robot-open-ended-autonomous-learning
[3] Differentiable Physics and Stable Modes for Tool-Use and Manipulation Planning
https://youtu.be/-L4tCIGXKBE
[4] MoveIt
https://moveit.ros.org
https://youtu.be/0og1SaZYtRc
[5] OpenAI Deep Reinforcement Learning Baselines
https://github.com/openai/baselines
[6] Touch sensors
https://rebrand.ly/TouchSensors
Introduction to Neural Networks or Advanced Neural Networks courses.
Python or C++ ( > 1 year).
In case this would not find enough interest for a team project, this project proposal would be also offered (in reduced/modified form)
Frequency | Weekday | Time | Format / Place | Period |
---|
Module | Course | Requirements | |
---|---|---|---|
39-M-Inf-GP Grundlagenprojekt Intelligente Systeme | Gruppenprojekt | Ungraded examination
|
Student information |
The binding module descriptions contain further information, including specifications on the "types of assignments" students need to complete. In cases where a module description mentions more than one kind of assignment, the respective member of the teaching staff will decide which task(s) they assign the students.