Second World Congress on Biomimetics, Nano-Bio & Artificial Muscles, 2004  

 Symposium:  Towards Artificial Rodents (Tentative)

Special session chaired by Jean-Arcady Meyer & Agnès Guillot

Organizer & Chair:        Dr. Jean-Arcady Meyer ( jean-arcady.meyer@lip6.fr)

                    AnimatLab/LIP6, 8 rue du capitaine Scott, 75015 Paris, France

            Co-Chair:     Agnès Guillot 

                                AnimatLab/LIP6, 8 rue du capitaine Scott, 75015 Paris, France    

 

1.                  Biological fundations

Neurobiologically based proposals for structuring navigation strategies for autonomous agents.

Sidney I Wiener and Angelo Arleo

Laboratoire de Physiologie de la Perception et de l’Action, CNRS - Collège de France, Paris, France

 

Neurobiological studies have revealed brain systems that are likely substrates for spatial orientation and navigation (that is, planning and executing displacements to a point not perceptible from the point of origin). Neuroanatomical studies have demonstrated the architecture and microcircuitry of multiple parallel and hierarchical pathways. Experimental lesions of the respective specific modules have shown deficits in particular types of cognitive processing (such as path integration, or mapping functions) indicating that these structures carry out complementary functions. For example the hippocampal system is implicated in employing configurations of environmental cues for self-localization - and is capable of reconstructing spatial representations in the presence of only a partial subset of the cues learned previously. In contrast, lesions to the caudate nucleus impair approaches to visible cues, but spare the capacity to employ configurational cues. These respective brain structures thus are vital for enabling particular characteristic strategies employing certain types of cues and signal processing.

The signal processing for these processes is revealed in neurophysiological recordings of neuronal activity in navigating rats. These show the elemental information, such as coding in 'place cells', selective for the position of the rat in its environment and 'head direction cells' selective for the cardinal orientation of the head. In addition to on-line processing, these systems are also involved in memorization and recall of this information.

On the basis of these experimental data, an integrative framework is put forward where these processes are organized hierarchically according to complexity and during extended experience, processing becomes more rapid and automatic and is successively transferred to structures specialized for this. Many tasks can thus solicit alternation between different systems according to the types of cues and processing at each stage of a multistep route. This provides an efficient system by reducing computational demands to simpler types of processing whenever possible.

 

2.                  Sensors

Towards a Robot Whisker System Modelled on the Rat Mystacial Vibrissae

Ben Mitchinson* , Tony Prescott*, Kevin Gurney*, and Peter Redgrave*, Chris Melhuish**, Tony Pipe**, Martin Pearson**, Ian Gilhespy**

* Adaptive Behaviour Research Group, University of Sheffield, UK

** Intelligent Autonomous Systems Laboratory, University of the West of England, UK

 

We will outline our current programme of research aimed at developing a multi-whisker sensory system modelled on the mystacial vibrissae of the laboratory rat. Neurobehavioural studies suggest that the rat whisker system is used in locomotion, maintenance of equilibrium, object recognition and surface discrimination. The behavioural cues discernible by the whiskers are likely to include location, proximity, relative velocity, size, and texture of nearby surfaces and objects. One striking characteristic of this system is the "whisking" movement employed during exploration behaviour, in which the vibrissae are swept back and forth across objects and surfaces in a synchronised wave at a rate of between 5 and 11 hertz. This whisking movement has been likened to a human running their fingertips over a surface, and it appears that the rat is able to exploit both the spatial and temporal characteristics of the pattern of deflections derived from whisking to glean information about characteristics such as surface texture. 

We will report the initial results of a three year project to develop an active, multi-whisker system capable of object detection and surface-discrimination.  We will also report on some pilot neurobehavioural studies of rat whisking in a simple navigation task.

 

3.                  Control architectures

The neural basis of spatial orientation in rats: Electrophysiology, computational modeling, and robotics

A. Arleo *, T. Degris **, C. Boucheny *, S. I. Wiener *

* Laboratoire de Physiologie de la Perception et de l’Action, CNRS - Collège de France, Paris, France

** AnimatLab/Laboratoire d’Informatique de Paris 6, Université Pierre et Marie Curie, Paris, France

 

Place and direction representations are crucial for most models elaborating flexible spatial behavior. This has led to the hypothesis that hippocampal place (HP) cells and head direction (HD) cells may constitute vital neural bases for cognitive processes underlying navigation of rats. Indeed, electrophysiological single-cell recordings suggest that HP neurons can encode the position of the rat in allocentric (world centered) coordinates, while HD neurons provide an allocentric representation of the orientation of the head of the rat projected onto the azimuthal plane. However, the paucity of experimental evidence at the intermediate levels (e.g., neural populations, macro-circuits, and system level) makes it difficult to explicitly relate the single-cell level (e.g., electrophysiological properties of HP and HD cells) to the behavioral level and hence understand how HP and HD cells enable spatial navigation.

Theoretical modeling and neuromimetic robotics may offer useful tools to systematically explore hypotheses concerning these relations. Since neuromimetic artifacts are simpler and more `experimentally transparent' than biological systems, they may permit new hypothesis to be examined at different hierarchical levels and lead to a better understanding of the mechanisms supporting cognitive behavior of animals. Conversely, a biologically-founded approach to model spatial cognition offers the attractive prospect of developing autonomous systems that emulate the navigation capabilities of animals and may lead to beneficial application in improving design of more powerful and adaptive robots.

We propose an integrative neuroscience and engineering approach involving electrophysiological experiments, computational modelling, and neuromimetic robotics for studying spatial orientation. Here, we present a prototypic case and focus on HD cells, first presenting a series of electrophysiological experiments addressing the following issue: How do actual thalamic HD cells update their directional representation on the basis of visual landmarks in the environment? Secondly, a computational model of HD cells based on a continuous attractor network of leaky integrate-and-fire spiking neurons will be presented. The weight distribution determining the intrinsic dynamics of the attractor (i.e., recurrent inhibitory and excitatory signals) has been established by means of a genetic algorithm. The model captures several electrophysiological properties of single HD cells and, most importantly, permits to investigate the state transition dynamics (e.g., during visual reorientation) at the level of a large population of formal HD cells. The experimental validation of the model has been done via a robotic implementation, which has produced an artificial HD system endowing a robot with allocentric spatial orientation capabilities.

 

Towards a robot control architecture modelled on the 'integrative core' of the rat brain

Mark Humphries, Tony Prescott, Kevin Gurney, and Peter Redgrave

Adaptive Behaviour Research Group, University of Sheffield, UK

 

We will outline our current programme of research aimed at developing a biomimetic 'integrative core' for robust and effective robot control modelled on key structures in the rat brain. Our general focus is on two specific brain systems that are hypothesized to play an important role in action selection-the basal ganglia (BG) and the reticular formation (RF). This paper will specifically report on the development of a new model of the reticular formation inspired, in part, by the classic RF model of Kilmer, McCulloch, and Blum, 1969.

We will report preliminary robotic experiments using this model that aim to show that this structure can support a form of decentralized action selection. Finally, we will discuss the possibility that the RF forms a key element of a layered control architecture within the rat brain, functioning to provide the neural substrate for action selection during basal ganglia development or following damage to the basal ganglia.

 

Using Rodent Hippocampal Models for Robotic Simultaneous Localisation and Mapping

Gordon Wyeth and Michael Milford

School of Information Technology and Electrical Engineering, University of Queensland 4072 Australia


To navigate successfully in a novel environment, a robot needs to be able to perform Simultaneous Localization And Mapping (SLAM).  In this paper we present a biologically inspired system called RatSLAM that uses visual input from a colour camera to carry out SLAM on an autonomous robot.

 The heart of the RatSLAM system is based on current integrative models of the rodent hippocampal complex. The RatSLAM system is shown to effectively perform parameter self-calibration and SLAM in one-dimension.  Tests in two dimensional environments reveal the inability of the RatSLAM system to maintain multiple pose hypotheses in the face of ambiguous visual input. 

These results support recent rodent experimentation that suggests that current models of the hippocampal complex are not the complete solution to SLAM problem.  

 

Navigation skills of an artificial mouse

Verena Hafner

Artificial Intelligence Laboratory, Department of Information Technology, University of Zurich, Switzerland

 

Mice and rats are probably the best studied vertebrates with respect to their navigation skills. The discovery of place cells in the rat's hippocampus in the early 1970s inspired the modelling of such cells - whose firing rate correlates with the spatial position of the animal - and the study of their importance for navigation behaviour. This work shows how such a model can be applied and optimised on a mobile robot which learns to navigate within its environment through exploration. The place cell model is enhanced with local navigation skills. The major sensory information is gained from a camera providing the robot with omnidirectional vision, similar to the large visual field of rats and mice. We are currently investigating how tactile information from an active whisker array on the robot can be used as an additional sensory modality for the place cell model.

 

Complex Cell Representations for Robot Localisation

Gordon Wyeth and David Prasser

School of Information Technology and Electrical Engineering, University of Queensland 4072 Australia

Various researchers have explored the hierarchical nature of the mammalian visual cortex for the purposes of building visual classifiers. In this paper we use a hierarchical model of the visual cortex to learn the localisation of an autonomous robot based on visual input. A key factor in the development of a biologically plausible vision system for localisation is the representations used in the complex cells of the visual cortex. The invariance properties of the complex cell representations are critical to successful robot localisation. 

The effectiveness of two different complex cell models is evaluated, and is shown to be suitable as the basis for localisation. Tests with a simple neural network show that reasonable localisation can be achieved with single-shot learning.

 

4.                  Learning

Actor-Critic models of reinforcement learning in the basal ganglia: from natural to artificial rats

M. Khamassi *,**, B. Girard*,**, A. Guillot*, and A. Berthoz**

* AnimatLab/Laboratoire d’Informatique de Paris 6, Université Pierre et Marie Curie, Paris, France

** Laboratoire de Physiologie de la Perception et de l’Action, CNRS - Collège de France, Paris, France

 

Actor-Critic architectures have been proposed as models of dopamine-like reinforcement learning mechanisms in the rat’s basal ganglia. In such models, an Actor network learns to select actions so as to maximize the weighted sum of future rewards, computed on line by another network, a Critic. The Critic predicts this sum by comparing its estimation of the reward with the actual one by means of a Temporal Difference (TD) learning rule, in which the error between two successive predictions is used to update the synaptic weights. A strong resemblance between the activation patterns of dopaminergic neurons that project to the basal ganglia, and the TD error prediction signal centers these models on the role of dopamine in the Critic. A recent review of numerous computational models, built on this principle since 1995, concluded on several issues raised by the inconsistency of the detailed implementation of Actor and Critic models with known basal ganglia anatomy and physiology. Here we consider some of the main issues, updated with anatomical and neurophysiological knowledge.

We also illustrate the consequences of different hypotheses concerning the Critic by comparing three existing models, connected with the same Actor, in a simulated robot as it performs the same reward-seeking task. During this task, the robot freely moves in a plus maze where a reward is distributed at different sites and referent to site-specific local stimuli. The robot has to select sequences of locomotor acts that enable it to reach the goal from any place in the maze. The differences between the simulated Critics lie in their capacity to build short or long sequences of stimulus-response associations to attain the reward, and on their capacity to construct them autonomously – i.e. with or without preprogrammed associations. The structure of these learned behavioral sequences is compared with the neurophysiological activation data of brain striatal cells recorded in real rats performing the same task.

A general discussion evokes the growing number of works that question the two main assumptions on which all Actor-Critic models are designed, i.e. the role of dopamine as a reinforcement signal, and the TD rule as a learning process for the basal ganglia.

The perspective of this work is to implement learning mechanisms processing in the basal ganglia in the control architecture of the artificial rat Psikharpax, already inspired by the same neural structures.  

 

Artificial Rodents that Can Predict the Future

Nestor A. Schmajuk

Department of Psychological and Brain Sciences, Duke University, U.S.A.

 

We describe a neural network model that characterizes a remarkable large number of classical conditioning paradigms. The network 

(a) describes behavior in real time,

(b) contains simple and configural stimulus representations, and

(c) includes attentional control of both storage and retrieval.

 

The behavioral properties of the model are supported by rigorous computer simulations, which are consistent with a large body of experimental data. The network can be applied to the design of predicting mechanisms that can be used in an artificial rodent.

 

5.                  Evolution

The Cyber Rodent Project

Eiji Uchibe, Genci Capi, Stefan Elfwing, Anders Eriksson, Hirofumi Suzuyama, Kenji Doya

ATR Computational Neuroscience Laboratories; CREST, JST, Japan

 

Successful acquisition of behaviors of autonomous robots by learning and evolution often requires appropriate selection of reward or fitness functions and meta-parameters of learning and evolution algorithms. Automatic regulation of these factors is not only necessary for building truly automatic robots, but also helpful for understanding the mechanisms of motivation and emotion in animals and humans.

In order to elucidate the functions of the reward system, we built a robotic platform, Cyber Rodents, which have the same vital constraints as biological agents, namely, self-preservation and self-reproduction.

Cyber Rodents are two-wheel driven rodent-like robots that can recharge itself from battery packs scattered in the field and can copy their ' genes' (programs or parameters) through IR communication ports.

We report our preliminary results of our simulations and experiments, such as evolution of metaparameters and hierarchical architectures.

 

6.                  Integrative projects

The Psikharpax project: towards building an artificial rat.  

Jean-Arcady Meyer *, Agnès Guillot *, Patrick Pirim **

* AnimatLab/Laboratoire d’Informatique de Paris 6, Université Pierre et Marie Curie, Paris, France

** BEV, Boulogne Billancourt, France

 

The Psikharpax project aims at endowing a robot with a sensory-motor equipment and a neural control architecture that will afford some of the adaptive capacities that are exhibited by real rats.

Concerning the sensory-motor equipment, the article first introduces the robot's visual system and focuses on the implementation of the vestibulo-ocular reflex, on the management of saccades, and on the mechanisms of landmark detection. Then it describes the robot's tactile (whiskers) and auditory (electronic cochlea) systems.

Concerning the control architecture, the paper specifies how localization and navigation capacities are implemented thanks to a biomimetic model of hippocampal and para-hippocampal structures in the rat brain. Likewise, it also reports how action selection capacities are afforded by another biomimetic model inspired by the anatomy and physiology of the basal ganglia. Finally, a model calling upon the cortical columns in the frontal cortex makes planning possible.

Preliminary results on the implementation of learning mechanisms in these structures are also presented.

Finally, the paper describes how Psikharpax will integrate these adaptive mechanisms and structures in a functional and coherent system that will be able to explore an unknown environment, to build a cognitive map of it, and to use this map to plan trajectories to places where the robot's will fulfil its various motivations, like "eating", "drinking" and avoiding "predators".

 

The integrative biomimetic architecture of Psikharpax : action  selection and navigation in realistic 3D environments

L. Lachèze*, B. Girard**, A. Guillot*, and J.-A. Meyer*

* AnimatLab/Laboratoire d’Informatique de Paris 6, Université Pierre et Marie Curie, Paris, France

** Laboratoire de Physiologie de la Perception et de l’Action, CNRS - Collège de France, Paris, France

  

A biomimetic control architecture inspired by the rat basal ganglia integrates action selection and navigation through the simulation of dorsal and ventral circuits. The input module of the ventral part corresponds to the nucleus accumbens (Nacc), which selects locomotion actions calling upon various navigation strategies and motivations. The input module of the dorsal part corresponds to the dorsal striatum, which is in charge of the selection of non-spatial tasks and of the coordination with Nacc's decisions. This architecture was previously implemented on a simulated robot that had to survive in an environment where places for `energy ingestion' and `energy digestion' could be found. The corresponding results demonstrated that the robot was able to ensure its survival with the selection of efficient sequences of actions, using its abilities of map-building and path-planning.

In this work, the same architecture is tested with another robot and another environment that are more complex than the previous ones. In particular, the physical characteristics of the robot are more realistic, as they closely emulate those of an existing robot. Moreover, the robot has to deal with a greater set of motivations and actions, and is equipped with more complex visual and orientation systems. Several experiments are performed in specific environments that serve to assess specific adaptive abilities. Other experiments involve a large office-like environment, cluttered with numerous obstacles, in which the full robot's adaptive capacities are put at work. Results demonstrate the capacity of the architecture to generate behavioral sequences that ensure survival. They also support the next step of the research, i.e. the embodiment of such a biomimetic model into a real robot and a real office.