Robot behaviour multi-level integrated learning method and robot behaviour multi-level integrated learning system

Abstract

The invention discloses a robot behaviour multi-level integrated learning method and a robot behaviour multi-level integrated learning system and relates to a robot behaviour control technology. The robot behaviour multi-level integrated learning system comprises a data sampling module, an unsupervised learning module, a supervised learning module, a motivation learning module and a command output module, wherein the data sampling module is used for inputting the information data of a robot; the unsupervised learning module acquires an environment mode characteristic vector which is used for representing the real-time change of an operating environment of the robot; the supervised learning module maps the environment mode characteristic vector into a motion command of a required behaviourof the robot in a real-time on-line mode by using the environment mode characteristic vector as an input signal; and the motivation learning module carries out real-time on-line optimization and accurate setting on parameters of a robot behaviour controller to make the controller operate reliably and stably. The method and the system provided by the invention can be applied to different types of robots which are capable of learning new robot behaviours and optimizing the conventional robot behaviours to adapt to the dynamic change of the operating environment, so that the intelligent and self-control capability of the robot is improved, the generality of the learning system is improved, and the design of the controller is simplified.

Description

technical field [0001] The invention relates to robot behavior control technology, in particular to a robot behavior learning system and method integrating multiple modes. Background technique [0002] There are already many engineering techniques for the design of robot behavior controllers, such as traditional AI based on symbolic reasoning, fuzzy logic, and mature linear control methods in cybernetics. However, these methods either require a definite environment model to realize "perception-plan-action", or require complex design and analysis of controller parameter tuning and system stability. Therefore, once the robot and its corresponding operating environment have any changes, the original planning strategy becomes outdated, and the designer has to redesign the controller, which increases the cost and cycle of system design. [0003] In recent years, there have also been some robot behavior learning systems based on neural networks that integrate more than two learni...

AI Technical Summary

Problems solved by technology

However, these learning systems have not been implemented on actual robots because of the long training time of incentive learning, or they only perform offline learning without integrating the important learning mode of "incentive learning".

Therefore, it cannot satisfy the adaptive learning of robot behavior in a dynamically changing environment.

Images