clamesc/epuck-box-pushing

TUM_ARL_SS16: With its proximity sensors the ePuck robot can estimate the relative position of a small box and learn to push it in a given direction using Reinforcement Learning.

ePuck Box-Pushing Project

Final presentation

This is the final code for the Applied RL class in SS2016 for the only ePuck group consisting of

Claus Meschede Jan Rudolph José Luís Gaytán

It contains a unmodified copy of the code from the LDV RobotAPI repository in the robot folder.

Collected Data

The Data folder contains training data collected during test runs. 2016_07_05-2 is the set of data used for the graphs in the final report.

• The Q.npy contains the history of Q values after every episode.
• The s.npy contains the number of steps needed for every episode.
• The T.npy contains the average state/action transition probabilities.

You can use the plot script to reproduce the QValues/learning curve graph from the final report. The average state/action transition graph is created manually using the matrix in T.

Main Scripts

In order to retrieve distance sensor calibration parameters for the boxtracking algorithm, you can use the boxtracking class from the boxtracking_calibration file. The proximity_values.txt in this zipfile was taken for the ePuck with the ip address 192.168.1.203.

The make_epuck_run script is the one we used for training.

The show script is the same script, the only difference is that it uses 100% greedy actions.