Zsolt Kira

Site Map | Home
Bio | CV | Schedule | Public Key
Statement | Heterogeneity & Knowledge Sharing | Multi-robot Systems | Single-Robot Learning
Software Released
All Albums | IROS & EpiRob 2007 | Burning Man '06 | Europe '05 | Japan '04 | Nantahala '03
Bio | subglobal8 link | subglobal8 link | subglobal8 link | subglobal8 link | subglobal8 link | subglobal8 link

Single-Robot Learning

I strongly believe in building adaptive robots that can learn through experiences. To that end, I have used various machine learning techniques, most notably including:

Unsupervised Learning for Fault Detection and Localization

Building sensor mappings to detect faults.
In this first work, we researched methods of maintaining flexible models of robot capabilities in order to detect changes or faults [1]. For example, we were able to detect faults such as a loose wheel on a Pioneer robot or a tilted camera. We have devised a complete framework that takes in information about what sensors the robot has and what sensor processing occurs, and automatically find stable mappings that occur while performing a task. The framework was implemented for our example using self-organizing maps to detect changes in the correlations between different types of sensing and issued motor commands.

The approach we took was one of learning correlations between sensing at multiple levels of abstraction during normal operation. For example, we correlate not just odometry and sonar data, but also optic flow data processed from a camera image. We also use unsupervised learning to model the correlations to avoid having to model individual fault types beforehand, since they might not be known.

Case-Based Reasoning for Behavioral Parameter Adaptation
ATRV-Jr in experiments conducted to evaluate original CBR learning system.

In another line of research, we used a different machine learning technique, Case-Based Reasoning, to automatically adapt parameters of low-level reactive robot behaviors through experiences [2]. Case-Based Reasoning is a method where previous problems or situations, in addition to their solutions, are remembered. When the agent encounters a similar situation again, it can adapt the previous solution to the current one and use it (accommodating for the differences between the two situations). The work done in this lab applied Case-Based Reasoning in order to learn which behavioral parameters to use in a behavior-based system.  As part of the MARS project, I added a memory management system to this existing framework whereby cases could be removed if a new situation arises that is not in the current library, but the library is full. Several metrics for case removal were compared, and it was shown that forgetting can significantly increase performance in some situations.

Below are multimedia related to the projects. You can scrolls through the pictures or start the movie with the arrows on the lower right.


©2009 Zsolt Kira