Original Data Format

zip

Name

Version mldata

Comment

Names

Data (first 10 data points)

ZIP archive

dataTrajectoryPrediction.txt, README

Description

See original paper http://user.cs.tu-berlin.de/~jetchev/10-nikolay-ICRA.pdf

The file "dataTrajectoryPrediction.txt" contains the features and targets we used for regression of trajectory costs. There are 20 000 data points, each is a single line consisting of the 297 feature dimensions and the last number (298th) is the log of the trajectory cost, the regression target.

Every 20 consecutive data points are part of 1 trial, so we have 1000 trials, each with 20 data points. In our paper, we used linear SVR to predict the costs for all points, and then for each of the 1000 trials found the trajectory with minimal predicted costs (out of the 20 choices), and then calculated the expected costs when making such choices. The reader is welcome to test his algorithms on this data set, and report his results to us.

REMARKS; - use crossvalidation and split in train/test sets. We trained on 750trials (15000 data points) and tested on 250 trials(5000 data points) - it ican be also interesting to predict the costs as a regression, but the real utility in robotics is choosing the best out of 20 possibilities, as said above. - Predicting the costs for all points and just looking at the prediction error is influenced by the preprocessing of the target costs, but selecting the best of 20 choices is independant of any transformations, as long as they are monotonous.

URLs

http://user.cs.tu-berlin.de/~jetchev/10-nikolay-ICRA.pdf

Publications

Data Source

Measurement Details

Usage Scenario

revision 1

by jetchev on 2010-11-25 13:55

revision 2

by phoyer on 2010-12-03 15:22

No one has posted any comments yet. Perhaps you would like to be the first?

Leave a comment

This item was downloaded 2704 times and viewed 1425 times.

No Tasks yet on dataset Robot Trajectory Cost Prediction