80 likes | 194 Views
System Identification (Ulrich Nehmzow). Random data Built ARMAX model worked well for different regression orders Trained RBF on data worked for ideal outputs, not for entire data Wallfollowing Computed correlations between 3 laser sensors and output (rot_speed):
E N D
System Identification (Ulrich Nehmzow) • Random data • Built ARMAX model worked well for different regression orders • Trained RBF on data worked for ideal outputs, not for entire data • Wallfollowing • Computed correlations between 3 laser sensors and output (rot_speed): 45° 90° 135° • Built ARMAX model to predict output from laser input; resulting coefficients corresponded to sensor correlation; Spearman rank .89 • Trained RBF; Spearman rank .50 to .91, depending on parameters
Imitation learning (Jan Peters) • Approaching red static object by steering Eddy
Imitation learning (Jan Peters) (2) Training to follow moving object by steering Eddy
Imitation learning (Jan Peters) Result: Action-state space
Imitation learning (Jan Peters) (3) Eddy, imitating object-following behaviour autonomously using learned regression model
Novelty Detection (Ulrich Nehmzow) • Build normality matrix and find specified outliers (2) Find one outlier in sensor data