Emergent Walking Behaviour in an Aibo Robot

by Cecilio Angulo, Ricardo A. Téllez and Diego E. Pardo

A totally distributed neuro-evolved architecture, designed for the general control of autonomous robots, has been employed for the emergence of walking behaviour in the 12 degrees of freedom (DOF) quadruped robotic Aibo dog. The concept of central pattern generators in animals is bio-inspired, and is used to obtain the desired walking robot. Moreover, the search to dissolve the 'Body-Mind' problem is our inspiration.

Our project is focused on the reactive interpretation of the mind, so we are interested in the emergence of certain behaviour. Quadrupedal walking is a complex task. In a robot like Aibo, there are twelve degrees of freedom, and their coordination to obtain a walking pattern becomes very difficult. To deal with this problem, the Knowledge Engineering Research Group (GREC) at the Technical University of Catalonia (UPC) has designed a neuro-evolved controller for the Aibo robot using neural networks distributed all over the robot. We have based our neuro-control system on Central Pattern Generators (CPGs). Biological CPGs are composed of groups of neurons that produce oscillatory signals without oscillatory inputs. It has been discovered that the walking movements of cats and dogs are governed by such elements, and it is thought that humans behave in the same way. We implement artificial CPGs using artificial neural networks (ANNs). To overcome the problem of capturing the dynamics of the system, we have used continuous-time recurrent neural networks (CTRNNs).

The proposed distributed neuro-evolved architecture is based on several uniform modules known as Intelligent Hardware Units (IHU). These can be designed around any physical device in the robot (sensor or actuator). Each IHU is composed of a sensor or an actuator and a micro-controller that implements an ANN. The ANN processes the information from its associated device (ie received sensor information or commands sent to the actuator). Figure 1 illustrates a simplified two sensor/two actuator architecture.

Figure 1
Figure 1: 'Me' depends on the 'task' (task-directed training) as interpreted by the translator; on the 'environment' ('outer world') interpreted by the modeller; and on the 'body' ('inner world', sensor and actuator), acting through the controller.

Through the use of a neuro-evolutionary algorithm, modules learn how to cooperate and how to control or interpret their associated elements, thereby allowing the whole robot to learn how to walk.
It should be stated that when several IHUs work together on a control task, each element has its own particular view of the situation, since each is in charge of a different sensor or actuator. This leads to a situation where each unit uses its knowledge both of the global situation and of its particular device to decide what its next action will be.

In this way, the traditional 'Mind-Body' problem can be dissolved. Control of the process is performed through three elements (modellers, controllers and translators). These elements must adapt the relationship between the 'body' (embodiment) and the 'environment' (situatedness).

The modeller is a control element that tries to adapt what the sensor 'sees' from the 'outer world' to what the cerebrum 'sees' in its 'inner world' by also considering the actions of the controller. Is it the nervous system? In any case, it undergoes continuous learning in order to adapt the body to the environment.

The translator is a control element that tries to translate the external set point (a behaviour associated to a task) as an interpretation of the 'outer world'. Is it the behavioural system? In any case, it is the learning function for the whole 'inner world' system.

The controller is a control element that deals with both, the internal perception of the 'outer world' (sensor) in the form of the 'inner world' units (modeller), and the task to be accomplished (external set point) translated to a internal set point, also in the 'inner world' units (translator); then appropriated commands are sent to its associated actuator. Is it the cerebral neural network? In any case, it must drive the actuator changing the body/environment situation.

Figure 2: Aibo walking sequence.
Figure 2: Aibo walking sequence.

Until now, embodied AI approaches have focused on lower-level aspects of the behaviour related to their embodiment; this is the reactive interpretation of the mind. However, our proposed distributed architecture is able to accomplish complex tasks, like the walking behaviour of the 12 DOF Aibo robot (Figure 2), by dissolving the 'mind-body' problem. The next stage in this project is to extend the architecture so as to be able to model higher-level behavioural aspects. In this way we avoid hybrid models that integrate both a 'deliberative' and an embodied 'reactive' component.


Please contact:
Cecilio Angulo-Bahón, Ricardo Agamenón Téllez-Lara, Diego Esteban Pardo-Ayala, Technical University of Catalonia (UPC) / SpaRCIM, Spain
Tel: +34 93 8967798
E-mail: cecilio.angulo@upc.edu, r_tellez@ouroboros.org, diego.pardo@upc.edu