Learning in Robotic Arm Simulator

Equipe: SequeL (Sequential Learning), Inria
Responsable HDR: Prof. Philippe Preux
Encadrant: Michal Valko
Contexte: Machine Learning (Artificial Intelligence)
Problématique: Learning from demonstrations

Description:

This project involves learning in a robotic system from the human demonstrations. The following video shows the demonstration of placing a ball in a cup, that were performed at MPI Tubingen and TU Darmstadt.

Ball in a Cup Demonstration
▶

Click to watch the demonstration on YouTube

Not all of the demonstrations are successful and not all demonstration maybe examined. There are also several ways how to successfully perform that maneuver. The goal of this project is to find out if we can learn from all possible demonstrations.

Plan of work:

The approach we will apply the approach called "semi-supervised apprenticeship learning". The work will involve modification of the simulator's code to perform this learning. The input data are the recorded trajectories in the above video. The data and the source code of the simulator will be provided. The aim is to apply/implement the semi-supervised algorithm and evaluate it performance on the "ball in a cup" task.

References

[1] Christian Daniel, Gerhard Neumann, Marc Deisenroth, Jan Peters
Robot Arm Evaluation Scenarios (Chapter 3)
CompLACS Deliverable
[2] Michal Valko, Mohammad Ghavamzadeh, Alessandro Lazaric
Semi-supervised inverse reinforcement learning
European Workshop on Reinforcement Learning (EWRL 2012)
[3] Abdeslam Boularias, Jens Kober, Jan Peters
Relative Entropy Inverse Reinforcement Learning
International Conference on Artificial Intelligence and Statistics (AISTATS 2011)
PDF
mv