Date, Time & Location
Sept 21st, 2022. 10.00-11.00. Virtual Meeting (Zoom).
Participants
- Prof. Dr. Elmar Rueckert
- Vedant Dave, M.Sc.
- Prof. Dr. Philipp Beckerle
- Ege Cansev, M.Sc.
- Adna Bliek, M.Sc.
Agenda
- Update on the Erlangen Team progress
- Update on the Leoben Team progress
- Next meeting
Details on the Meeting Notes
TOP 1: Update on the Erlangen Team progress by Philipp
Adna’s master student that worked on RL submitted his/her thesis. In this work the robot arm learned point-to-point movement.
Ege’s glove design is about to be finished soon. One more flex sensor is added to the thumb to better capture fine thumb motions.
- Experiment (Atari): The goal is to evaluate learning algos by humans. What are good quality criteria for humans? The idea is to show humans different videos of the performing algorithms. Evaluated will be human preferences and trust.
- Experiment (Robot, wire-maze): Goal: humans teach the robot via kinesthetic-teaching or teleoperation. Thereafter the robot’s performance is evaluated (per video for example). Posible options include transfer learning tasks, where the maze is changed or shared control where the user corrects pre-trained skills.
Idea: Learn initially a pre-trained policy (e.g., a neural net in simulations) that does not use all joints. Transfer the task to involve, e.g., curves which require more complex skills. The user corrects the robots policy and the model is updated. The error signal used for updating the model is the difference between the model output and the user correction (e.g., in the x,y Cartesian space). The alternative is to record user demonstrations of successful behavior and to retrain the model.
TOP 2: Update on the Leoben Team progress by Elmar
Niko will work on ROS-2.
Fotios is working on turning his contrastiv learning and vae for classification. He is currently extending his work to regression tasks.
Vedant is working on inverse dynamics model learning. For the project this could be a second experiment. Ege suggested to collaborate on the shared weight lifting experiment that Erlangen plans to investigate.
TOP 3: Next meeting November 9th 2022 14.00-15.00
Next meeting is scheduled for the date above using ZOOM.
Adna will initiate monthly PhD meetings.