Sensory-motor Coordination
Table of Contents
Description
Cognitive processes at high abstraction levels rely on a hierarchy of lower level behaviors. Low level autonomous behaviors can be constructed from basic sensorimotor behaviors. A basic sensorimotor behavior is a reflex, which is a direct motor response to sensory stimuli. Basic behaviors are fundamental to a behavior based robot. If basic behaviors are determined a priori by robot designers, the final controller is inefficient at best, or completely wrong at worst. An alternative to overcome these limitations is to acquire basic behaviors by the robot itself.
Because of the experimental accessibility, the plant's simplicity, and the diverse collateral knowledge about the visual system, gaze control is the best-studied control system. Gaze control skills are between instinct and experienced, they provide a promising domain for the investigation of control theory, algorithms and neural mechanisms underlying the acquisition and performance of general sensorimotor skills.
In a long run, I want to study how to construct non-linear temporal sensory-motor mapping so the robot can emerge more complex behaviors from its interaction with its working environment. It is closely related to reinforcement learning and imitation learning.
Publications
- Chenggang Liu; Jianbo Su, "Basic behavior acquisition based on multisensor integration of a robot head," Robotics and Automation, 2008. ICRA 2008. IEEE International Conference on , vol., no., pp.3094-3099, 19-23 May 2008 URL: http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=4543681&isnumber=4543169 Liu, C., & Su, J. (2009).
- Basic Behavior Acquisition with Multisensor Integration for Robot System. Advanced Robotics, 23(12-13), 1-17.
Demos
The video demonstrates how the 'basic behavior acquisition based on multisensor integration' works. First, the robot moves its head randomly to build up the mapping between its motor commands and the changes of features. Then the user specifies a desired state in the feature space. Having the sensory-motor mapping, the robot could figure out right motor commands to achieve the desired state in the features space. The benefits include: behavior control without specific programming automatic multi-sensor integration to accomplish a task In this experiment, the mapping is a static linear mapping, which is estimated by a Kalman filter.
This video demonstrates the user interface for the Basic behavior acquisition with multisensor integration. The three windows on the top (from left to right) are: 1) face detection 2) features for optical flow tracking and 3) retinal image.