Hung Ngo

HRI application demo

Efficient Multiclass Prediction from Binary Feedback

Short version demo video with voiceover will appear at HRI 2014 Video Demon Session:

In this demo the robot learns to recogize hand gestures from a human instructor while executing some tasks. The instructor just gives binary feedback (right/wrong) for each prediction of the robot. To learn from such limited feedback, the robot uses Upper Confidence Weighted Learning (UCWL) algorithm as described in our publications:

Longer version of the demo video without voiceover:

Video description:

At the beginning of each interaction, the robot blinks its beacon. This signals to the instructor to start the interaction session. The instructor then shows the robot a hand gesture, one out of a set of six (finger counts). Afterwards, the robot blinks its LEDs with a color encoding the predicted class (one to six, where six is a ``fist'' gesture). The instructor then provides the robot with feedback of right or wrong using two special gestures. He hides the glove to indicate a correct prediction, or he waves his fist vigorously to indicate an incorrect prediction.

If the robot's prediction was correct, it goes on to execute its task - in this case to go to the pole with the same color as the LED. Otherwise, it blinks its beacon again to signal the instructor to start a new session (i.e., to repeat the previous gesture).

The Footbot uses its onboard camera to capture an image of the hand gesture. It performs preprocessing on the image, i.e., color-based segmentation and normalization, etc. It then extracts features using a pretrained Convolutional Neural Network. The feature vector resulting from each interaction is then used to predict the gesture class, which also encodes the task that is requested of the robot. After the robot gets the binary feedback, it updates its online multiclass classifier using Upper Confidence Weighted Learning. In this demo, the classifiers are pretrained from only 30 training examples (roughly 5 per class).
Other Related Demo from IDSIA:

Last modified January 30, 2014.