Autonomous Vehicle Architecture Inspired by

the Neurocognition of Human Driving

Mauro Da Lio

, Alice Plebe

, Daniele Bortoluzzi

, Gastone Pietro Rosati Papini

and Riccardo Don

Department of Industrial Engineering, University of Trento, Trento, Italy

Department of Information Engineering and Computer Science, University of Trento, Trento, Italy

Keywords:

Automated Driving, Cognitive Systems.

Abstract:

The realization of Autonomous vehicles is recognized as a relevant objective for the modern society and con-

stitutes a challenge which in the last decade is concentrating a growing interest, involving both manufacturers

and research institutes. The standard approach to the realization of automated driving agents is based on a

well-known paradigm, consisting of the sense-think-act scheme. Even though this implements an understan-

dable and agreeable logic, a driving agent based on such an approach needs to be tested and qualiﬁed at a level

of reliability which requires a huge experimental campaign. In this position paper the scope of the problem

of automated driving is widened into the cognitive sciences, where the inspiration is taken to reformulate the

underlying paradigm of the automated agent architecture. In the framework of the EU Horizon 2020 Dre-

ams4Cars Research and Innovation Action project the challenge is to design and train an automated driving

agent which mimics the known human cognitive architecture and as such is able to learn from signiﬁcant

situations encountered (either simulated or experienced), rather than simply applying a set of ﬁxed rules.

1 INTRODUCTION

The recent years witnessed an exponential increase in

the number of industrial –and research– initiatives ai-

med at the implementation of automated driving. Ho-

wever, after the initial enthusiasms, one challenge in

particular is emerging: how to prove that the enginee-

red vehicles are safer than humans. One recent study

(Kalra and Paddock, 2016) demonstrates that exten-

sive road testing would require driving for billions

of miles –which is not realistic– to provide statisti-

cally signiﬁcant proofs that the software is safer than

humans. By comparison, companies testing Autono-

mous Driving (AD) vehicles and reporting to the Ca-

lifornia State have accrued less than 0.5 million miles

between 9/2014 and 11/2015 (Dixit et al., 2016). The

problem stems from the fact that humans –contrary to

superﬁcial perception– are very reliable at driving: in

the US there are about 33,000 fatalities and 2.3 mil-

lion injuries per year. As large as these ﬁgures may

look like, when divided by the total traveled miles (3

trillions) they correspond to very low rates of 1.09 fa-

talities and 77 injuries per 100,000,000 miles.

This position paper proposes a paradigm shift for

the architecture of the driving agents. Today, the soft-

ware for AD follows the sense-think-act paradigm,

which means that the perception system produces a

symbolic representation of the environment and that

there is software determining the agent behavior ba-

sed on this representation (a.k.a. Cartesian rationa-

lism). This approach almost necessarily means that

the software must be entirely developed and tested by

human designers. Simulations may be used to reduce

the amount of testing but the tested scenarios, again,

have to be known. For the sense-think-act architec-

ture, the goal of designing and testing a system ca-

pable of error-free behaviors for billions of miles is

deﬁnitely a challenge.

Overall, designing a system capable of correct be-

havior for billions of miles means designing a system

capable of acting in autonomy, i.e., in situations that

might have not been expressly considered at the de-

sign stage. Humans have the ability to transfer kno-

wledge of the few generic situations that they faced

in a few tenths of training hours at driving schools to

deal with novel situations found in a lifetime of dri-

ving. Moreover they continuously learn: for example,

every critical situation (even only potential hazards)

is detected and imagery mechanisms to study that si-

tuations at both dreaming and wake states exist, that

allow humans to learn new behavioral strategies for

situations that they have only imagined. That is why

Da Lio, M., Plebe, A., Bortoluzzi, D., Rosati Papini, G. and Donà, R.

Autonomous Vehicle Architecture Inspired by the Neurocognition of Human Driving.

DOI: 10.5220/0006785605070513

In Proceedings of the 4th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2018), pages 507-513

ISBN: 978-989-758-293-6

507

young licensees have far more risk than senior drivers.

Human accidents are largely due to distractions,

tiredness, risk taking or driving under the effects of

alcohol or drugs, aspects that will not affect artiﬁcial

systems. Hence we want to take inspirations from the

mechanisms underlying human autonomy and pro-

pose an artiﬁcial cognitive architecture capable of si-

milar autonomy, and capable of increasing its abilities

like senior drivers. For this, we analyze in the rest of

the paper the human sensorimotor system and show

how it might be taken as a model for the architecture

of driving agents.

This paper is one output of the EU Horizon

2020 ”Dreams4Cars” Research and Innovation action

(www.dreams4cars.eu).

2 THE NEUROCOGNITIVE

UNDERPINNING OF DRIVING

The ability to drive is one the many highly specialized

human sensorimotor behaviors. In addition to the uni-

versal ability to walk, run, and jump, specializations

that share with driving complex motor coordination

include playing piano, skying, playing tennis, and the

like (Wolpert et al., 2011). The history of robotics

has proved how difﬁcult it is to implement perception-

action systems that lead to performances similar to

human ones, even for the basic walking abilities. Au-

tomated driving may appear easier, since car motion is

more prone to direct computer control than, for exam-

ple, robots with human-like gait. However, we deem

that the sophisticated control system that the human

brain develops, when learning to drive by comman-

ding the ordinary car interfaces, steering wheel and

pedals, may reveal precious insights on how to imple-

ment a robust automatic driving system.

The task to drive is solved by the human brain with

the same kind of strategy that we adopt for every sort

of motor planning that requires continuous and com-

plex perceptual feedback. For sure, how this general

strategy works is far from being fully understood, and

there are several competing theories, lacking a unani-

mous consensus. Nevertheless, a huge body of rese-

arch in neuroscience and cognitive neuroscience has

been produced in the past decades, that allows us to

grasp some general cues useful for designing artiﬁcial

car control systems.

2.1 Simulation and Emulation

Primarily, we consider approaches that belong to the

simulative theories, in the account reviewed in (Hes-

slow, 2012). In cognition, the idea of simulation has

been traditionally related with the hypothesis that our

social cognition is based on “simulating” internally

what is going on in the mind of others, adopting a

sort of psychological theory of how (ours and other’s)

minds work (Gallagher, 2007). The simulation the-

ory supported by Hesslow points to a different di-

rection, he argues that thinking in general is explica-

ted by simulating perceptions and actions involved in

the thought, without the need of actually executing the

actions, or perceiving online what is imagined. This

is also the course of action our project intends to take

forward (Da Lio et al., 2017). In the view of Hesslow,

simulation is a general principle of cognition, expli-

cated in at least three different components: actions,

perception, and anticipation.

One of the earlier proposals in this direction is the

theory of neural emulators (Grush, 2004), that bridges

a close link between the engineering domain of con-

trol theory and signal processing to neural represen-

tations. According to the emulation theory, the brain

is able to constructs neural models of the body and

of the environment, in addition to simply engaging

with them. During overt sensorimotor engagement,

these models are driven by efference copies in paral-

lel with the body and the environment, so to generate

expectations of the sensory feedback. Later on, these

models can also be run ofﬂine, in order to predict out-

comes of different actions, and evaluate and develop

motor plans. In terms of control theory, Grush’s emu-

lators are essentially forward models, and so will be

called in the rest of the paper. Grush further argued

that forward models can be realized by Kalman ﬁlters,

pointing to models that seem to support his hypothesis

(Wolpert and Kawato, 1998), and neuroscientiﬁc evi-

dences for forward models in the cerebellum (Wolpert

et al., 1998; Jeannerod and Frak, 1999). We will not

use the Kalman ﬁlter hypothesis here, while we em-

brace the characterization of the cerebellum as site of

forward models predicting the sensory effects of mo-

vements.

2.2 Imagery and Dreaming

A second tradition of research that converges into

Hesslow’s simulation account is perceptual imagery,

especially of visual modality. Mental imagery is the

phenomenon where a representation of the type crea-

ted during the initial phases of perception is present,

but the stimulus is not actually being perceived; such

representations preserve the perceptible properties of

the stimulus (Moulton and Kosslyn, 2009). Mental

imagery has connections with the phenomenon of dre-

aming, and it has been argued (Thill and Svensson,

2011) that the function itself of dreaming, in infants

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

508

through to early childhood, is to trigger and exercise

the capacity of simulating.

What is the mechanism that allows primary sen-

sorial areas to be activated both by online sti-

muli or by perceptual imagery remains to be unvei-

led. A prominent proposal is formulated in term

of convergence-divergence zones (CDZs) (Meyer and

Damasio, 2009). CDZs receive convergent projecti-

ons from the early sensorimotor sites and send back

divergent projections to the same sites. This arrange-

ment has the ﬁrst purpose to record the combinatorial

organization of the knowledge fragments coded in the

early cortices, together with the coding of how those

fragments must be combined to represent an object

comprehensively. CDZ records are built through ex-

perience, by interacting with objects. The CDZ fra-

mework can explain perceptual imagery, as it propo-

ses that similar neural networks are activated when

objects or events are processed in perceptual terms

and when they are recalled from memory. This pro-

ject will experiment with implementations loosely in-

spired by CDZs.

2.3 Main Neural Components of

Simulation

The simulation theory as spelled out by Hesslow co-

vers many aspects of human cognition, but does not

address how the simulation framework might be ac-

tually realized in the brain. This crucial aspect is,

instead, central in the investigations done by Cisek

(Cisek, 2007; Cisek and Kalaska, 2010). In his ge-

neral framework, simulation is one component of

the mechanisms by which the brain selects actions

and speciﬁes the parameters or metrics of those acti-

ons. Cisek’s theory is named affordance competition

hypothesis, in which “affordance” is the term origi-

nally used by Gibson (Gibson, 1979) to refer to the

action possibilities of the environment that are availa-

ble to an animal. For example, for most human-beings

a chair “affords” sitting, a glass “affords” grasping;

and we might say that for a driver a lane “affords”

either lane change or lane following and a yield line

“affords” either stopping or crossing. In Cisek, “af-

fordances” are the internal representations of the po-

tential actions which are in constant competition for

deciding the next behavior. In this project, the term

affordances will be used in exactly the same sense.

The ﬁrst neural component of the affordance com-

petition hypothesis, in the context of visually guided

actions, is the occipito-parietal dorsal stream. In the

traditional division of the visual processing path into

the dorsal stream and the occipito-temporal ventral

stream the former builds a representation of where

things are and the latter of what things are (Unger-

leider and Mishkin, 1982). A more recent account

of the dorsal stream proposes that its role is to medi-

ate various visually guided actions by several substre-

ams. For example, the lateralintraparietal (LIP) area

specify potential saccade targets; the medial intrapa-

rietal (MIP) area specify possible directions for arm

reaching.

The next fundamental component is made up by

the basal ganglia in connection with the dorsolateral

prefrontal cortex (DLPFC), and performs action se-

lection. Since action selection is a fundamental pro-

blem faced by even the most primitive of vertebra-

tes, it is consistent with the involvement of an ancient

structure conserved throughout evolution, like basal

ganglia. DLPFC appears to play the role of collecting

“votes” for categorically selecting one action over ot-

hers. There is a wide literature on the detailed me-

chanisms by which basal ganglia and prefrontal cor-

tex interact in taking decisions (Redgrave et al., 1999;

Doya, 1999; Bogacz and Gurney, 2007; Lewis et al.,

2011).

The last fundamental component is the cerebel-

lum, where an internal predictive feedback is gene-

rated, once the ﬁnal selected action is released. As

mentioned above, the cerebellum was also taken into

account by Grush for its theory of emulators.

3 AGENT SENSORIMOTOR

ARCHITECTURE

The overall architecture of the agent to be develo-

ped in this project is broadly derived from the neu-

ral components and brain strategies that allow humans

to drive, and were described in the previous sections.

For this reason, the main scheme of the agent, shown

in Fig. 1, is overlaid on a sketched brain, and is an

adaptation to the task of driving by an artiﬁcial agent

of the action-selection scheme in (Cisek and Kalaska,

2010, Fig.1 p.278).

In the remaining of this section the three main pro-

cessing pathways shown Fig. 1 will be described.

3.1 Inverse Model

The pathway depicted in red in Fig. 1 performs essen-

tially the transformation of an inverse model, taking

as input sensory data and producing outputs in the for-

mat of motor controls. This is the role played by the

visual dorsal stream in the account of Cisek, and we

adopt his terminology by calling affordances the ﬁnal

outputs of this processing pathway, (c) in the ﬁgure.

Autonomous Vehicle Architecture Inspired by the Neurocognition of Human Driving

509

Automotive

sensors

Motor Output

Cerebellum

(dorsal stream)

Action Selection

(basal ganglia)

motor

cortex

sensory

cortex

Figure 1: Main scheme of the agent architecture, the main pathways are depicted in red (dorsal stream), green (action se-

lection) and violet (forward model); (a), (b), and (c) are progressive stages of representations in a scale from sensorial to

affordances.

The format in (a) is basically in the topographic re-

presentation of the perceptual space, fusing informa-

tion from sensors like camera and LIDAR. The format

in (b) will be in a topology more oriented toward the

egocentric control requirements. This topology will

be in two dimensions, since the control space for dri-

ving can be reduced to the instantaneous lateral and

longitudinal controls. The format in (c) is still in a

two dimensional topology, but in terms of discrete af-

fordances, such as “lane change”, “car follow”, and

similar. As visible in Fig. 1, the architecture is arran-

ged to receive also input from intelligent sensors that

can provide data at higher levels of format, in (b) and

even in (c).

This pathway, much like the biological dorsal

stream, is in fact a collection of several specialized

substreams, at least the following can be identiﬁed:

• road geometry;

• obstacles;

• trafﬁc lights or other trafﬁc directives.

There is a difference in kind among the outputs of

those substreams: road geometry provides the active

regions in the space of longitudinal and lateral con-

trols, while obstacles and trafﬁc lights compute inhi-

bited regions.

At the core of the transformations from (a) to (c)

there will be a deep neural network model, loosely

inspired by CDZs (Meyer and Damasio, 2009) des-

cribed in §2.2. The deep neural network will encode

the high dimensional sensorial input data into layers

with gradually reduced dimensions, up to a ﬁnal low

dimensional feature space, reminding the convergent

process in Meyer & Damasio’s CDZs. This small fea-

ture space, in turn, becomes the input of a similar neu-

ral network acting as decoding, in which the dimensi-

ons of the layers are gradually increasing, paralleling

the divergence in CDZs. Note that, when the desired

output of the decoding network is the reconstruction

of the original input, the combination is an autoenco-

der, Hinton’s idea that gave birth to deep learning it-

self, before being called “deep learning” (Hinton and

Salakhutdinov, 2006; Hinton et al., 2006). Autoenco-

der has then become one of the most effective way of

generating new data with deep learning (Dosovitskiy

et al., 2017). In our case there will be both a decoding

network reconstructing perceptual data, which plays

the role of imagery, and a different network which de-

codes in the format of motor controls.

3.2 Action Selection

This component corresponds roughly to the proces-

ses performed in the brain by basal ganglia, typically

in coordination with DLPFC, as described in §2.3. It

has the function of deciding the agent’s behavior, and

can accept high-level biases, such as automation level

or driving style. Its implementation will be based on

the multihypothesis sequential probability ratio test,

an asymptotically optimal statistical test for decision

making that has been shown to be a possible computa-

tional abstraction of the decision function performed

by the cortex and basal ganglia.

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

510

3.3 Forward Model

This last component performs a role that is suppo-

sed to take place in the cerebellum, as discussed in

§2.3: forward model. The input of this component is

made up of efferent copies of motor commands, and

the output is in term of intermediate perceptual data

of the scene, as realized if the motor commands have

been issued. This way, it would be possible to detect

salient, and potentially dangerous situations, without

having actually experience those situations in reality.

4 SYSTEM ARCHITECTURE

The system here proposed implements three different

environments: the “wake” state, the “dream” state and

quality assurance. Fig 2 shows an outline of this ar-

chitecture.

4.1 The Wake State

This state corresponds with real driving, during

which the agent records situations that are conside-

red worthy to be re-enacted in the dream state. Those

salient situations are determined according to several

criteria. The most relevant events occur when there is

a discrepancy between the predictions of the internal

model and what actually happens in the real world.

These situations can point out some imperfections in

the prediction/planning model, or are the indication of

the occurrence of a novel condition the agent has no

knowledge about. Other relevant criteria that are ta-

ken into account are space-time separation with other

vehicles, jerkiness of control, compliance with trafﬁc

rules, and also trafﬁc and energy efﬁciency.

In order to be able to re-run experienced situati-

ons, the agent notes any event considered worthy of

further analysis. The agent logs both the low-level

sensory and control signals, and the high-level sig-

nals represented by the internal states of the agent’s

architecture. In such way, it is possible to record the

“intentions” of the agent (or the estimated intentions

of the human driver) at all levels of the sensorimotor

system. Hence, the agent is able to run simulations of

alternative lower-level strategies while preserving the

higher-level intentions.

4.2 The Dream State

During this state, the agent can explore a simula-

ted world to learning new behaviors, by testing dif-

ferent situations. The state is implemented in the

open-source open-source virtual reality driving simu-

lation environment OpenDS (www.opends.eu), based

primarily on multibody systems (MBS) technology

(Blundell and Harty, 2004), with some machine lear-

ning sub-model extensions. MBS is a general way for

engineers to simulate large-scale physical system dy-

namics, including complex systems of bodies under

the action of external forces, control loops, and other

conditions. One signiﬁcant advantage of MBS, com-

pared to e.g., cerebellar forward models, is longer-

term accuracy deriving from the universal physical

principles substrate.

MBS can be instantiated with a very large number

of parameters representing physical quantities. Se-

lecting the model parameters from a distribution can

be a method for generating imaginary conditions. Ne-

vertheless, the distribution values must be plausible (it

does not make sense to simulate a road with a friction

coefﬁcient equal to 10 since this never happens in re-

ality). Thus, the distributions can either derive from

the agent’s observations whilst in the “wake” state or

from a-priori knowledge, when it is is possible wit-

hout loss of generality.

A radical different approach for generating imagi-

nary situations is to use the same deep neural network

model at the core of the inverse model of the agent,

described in §3.1. This model has one principal de-

coding part that generates detailed motor commands

from features at low dimensions, but also a secon-

dary decoding part which works exactly as an autoen-

coder. It is therefore possible to generate a number

of imaginary situations by randomly exploring this

small feature space, with the caution of constraining

these random combinations towards those situations

that are more likely to be useful.

Another possible method is via generalized, top-

down “motor babbling”, a well established notion in

the framework of perception-action machine learning

(Shevchenko et al., 2009; Windridge et al., 2012;

Windridge, 2017) and applied in cognitive robotics

(Dearden and Demiris, 2005). It is akin to the pro-

cess of language learning in infant humans: starting

from minimal bootstrap sets of percepts and actions,

the entire action space is randomly sampled. For each

new motor action that produces a discernible percep-

tual output in the current perceptual set, the produced

percept is allocated. This process can be carried out

at various levels in the perception-action hierarchy.

4.2.1 Optimization

The ﬁnal purpose of the simulation environment is

the development of optimized behaviors (at all le-

vels). One of the possible strategy is optimal control

(OC), a methodology for producing optimal solutions

Autonomous Vehicle Architecture Inspired by the Neurocognition of Human Driving

511

Real Driving Quality Assurance

Simulation

(efﬁcient exploration, action discovery, optimisation)

Agent copy

Transfer salient

situations

Updated Agent

Certiﬁed Agent

(current)

Figure 2: The system architecture, consisting of three components: real driving, simulation/optimization and quality assu-

rance.

for goal-directed problems (Bertolazzi et al., 2005).

Despite being OC a methodological framework based

on differential algebraic equations and other mathe-

matical structures, there are empirical evidences that

the planning of hand movement to reach a target, as

performed by the brain, corresponds to the minimiza-

tion of the integral of the squared jerk, as in OC (Liu

and Todorov, 2007). Therefore, OC too is in line with

the brain-inspired design principle of this project.

Another possible approach is motivated learning,

inspired by the learning mechanisms in animals, and

thus better suited for higher-level strategies. In this

case, the agent takes advantage of opportunistic inte-

ractions with the environment to develop a knowledge

of what actions cause predictable effects in the envi-

ronment. In this way, the agent is able to build internal

models of action-outcome pairings in the brain.

4.3 Quality Assurance

Lastly, the third environment shown in Fig. 2 has

the purpose of quality assurance. As the previous

“dream” state, this environment is based on a multi-

body simulation system too, this one called CarMaker

by IPG Automotive (ipg-automotive.com).

By testing against several test cases, the system

certiﬁes new versions of the agent, ensuring that any

updated agent works no worse than the previously op-

timal one. It also helps to identify over-ﬁtting, such

as when the agent learns to better cope with the most

recently-dreamed situation at the expenses of earlier

ones.

5 CONCLUSION

This position paper has introduced a biologically

inspired layered control architecture for intelligent

vehicles, which relies on the recent developments in

cognitive systems. We believe that the adoption of

hierarchical perception-action learning via a dream si-

mulation mechanism can considerably extend the uti-

lity of training data during learning. Moreover, this

approach seems to offer advantages in terms of re-

sources needed for development.

ACKNOWLEDGEMENTS

This work is supported by the European Commission

under Grant 731593 (Dreams4Cars).

REFERENCES

Bertolazzi, E., Biral, F., and Da Lio, M. (2005). Symbolic-

numeric indirect method for solving optimal cont-

rol problems for large multibody systems: The time-

optimal racing vehicle example. Multibody System

Dynamics, 13(2):233–252.

Blundell, M. and Harty, D. (2004). The multibody systems

approach to vehicle dynamics. Elsevier, Amsterdam.

Bogacz, R. and Gurney, K. (2007). The basal ganglia and

cortex implement optimal decision making between

alternative actions. Neural computation, 19(2):442–

477.

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

512

Cisek, P. (2007). Cortical mechanisms of action selection:

the affordance competition hypothesis. Philosophical

Transactions of the Royal Society B: Biological Scien-

ces, 362(1485):1585–1599.

Cisek, P. and Kalaska, J. (2010). Neural mechanisms for

interacting with a world full of action choices. Annual

review of neuroscience, 33:269–298.

Da Lio, M., Thill, S., Svensson, H., Gurney, K., Anderson,

S., Windridge, D., Yksel, M., Saroldi, A., Andreone,

L., and Heich, H.-J. (2017). Exploiting Dream-Like

Simulation Mechanisms to Develop Safer Agents for

Automated Driving. Yokohama, Japan.

Dearden, A. and Demiris, Y. (2005). Learning forward mo-

dels for robots. International Joint Conference on Ar-

tiﬁcial Intelligence, 19:2–7.

Dixit, V. V., Chand, S., and Nair, D. J. (2016). Autono-

mous Vehicles: Disengagements, Accidents and Re-

action Times. PLOS ONE, 11(12):e0168054.

Dosovitskiy, A., Springenberg, J. T., Tatarchenko, M., and

Brox, T. (2017). Learning to generate chairs, ta-

bles and cars with convolutional networks. IEEE

Transactions on Pattern Analysis and Machine Intel-

ligence, 39(4):692–705.

Doya, K. (1999). What are the computations of the cerebel-

lum, the basal ganglia and the cerebral cortex? Neural

Networks, 12(78):961–974.

Gallagher, S. (2007). Social cognition and social robots.

Pragmatics & Cognition, 15(3):435–453.

Gibson, J. J. (1979). The Ecological Approach to Percep-

tion. Houghton Miﬂin, Boston (MA).

Grush, R. (2004). The emulation theory of representation:

motor control, imagery, and perception. The Beha-

vioral and brain sciences, 27(3):377–96; discussion

396–442.

Hesslow, G. (2012). The current status of the simulation

theory of cognition. Brain research, 1428:71–9.

Hinton, G. E., Osindero, S., and Teh, Y.-W. (2006). A fast

learning algorithm for deep belief nets. Neural Com-

putation, 18:1527–1554.

Hinton, G. E. and Salakhutdinov, R. R. (2006). Redu-

cing the dimensionality of data with neural networks.

Science, 28:504–507.

Jeannerod, M. and Frak, V. (1999). Mental imaging of mo-

tor activity in humans. Current Opinion in Neurobio-

logy, 9(6):735–739.

Kalra, N. and Paddock, S. M. (2016). Driving to safety:

How many miles of driving would it take to demon-

strate autonomous vehicle reliability? Transportation

Research Part A: Policy and Practice, 94:182–193.

Lewis, J., Chambers, J. M., Redgrave, P., and Gurney,

K. N. (2011). A computational model of intercon-

nected basal ganglia-thalamocortical loops for goal di-

rected action sequences. In BMC Neuroscience, vo-

lume 12(Suppl 1), page 136, Stockholm.

Liu, D. and Todorov, E. (2007). Evidence for the ﬂexible

sensorimotor strategies predicted by optimal feedback

control. The Journal of neuroscience : the ofﬁcial

journal of the Society for Neuroscience, 27(35):9354–

68.

Meyer, K. and Damasio, A. (2009). Convergence and di-

vergence in a neural architecture for recognition and

memory. Trends in Neurosciences, 32(7):376–382.

Moulton, S. T. and Kosslyn, S. M. (2009). Imagining pre-

dictions: mental imagery as mental emulation. Phi-

losophical Transactions of the Royal Society B: Bio-

logical Sciences, 364(1521):1273–1280.

Redgrave, P., Prescott, T., and Gurney, K. (1999). The ba-

sal ganglia: a vertebrate solution to the selection pro-

blem? Neuroscience, 89:1009–1023.

Shevchenko, M., Windridge, D., and Kittler, J. (2009). A

linear-complexity reparameterisation strategy for the

hierarchical bootstrapping of capabilities within per-

ceptionaction architectures. Image and Vision Com-

puting, 27(11):1702–1714.

Thill, S. and Svensson, H. (2011). The inception of simu-

lation: a hypothesis for the role of dreams in young

children. In Carlson, L., Hoelscher, C., and Shipley,

T. F., editors, Proceedings of the 33rd Annual Con-

ference of the Cognitive Science Society, pages 231–

236, Austin, TX. Cognitive Science Society.

Ungerleider, L. and Mishkin, M. (1982). Two cortical visual

systems. In Ingle, D. J., Goodale, M. A., and Mans-

ﬁeld, R. J. W., editors, Analysis of visual behavior,

pages 549–586. MIT Press, Cambridge (MA).

Windridge, D. (2017). Emergent Intentionality in

Perception-Action Subsumption Hierarchies. Fron-

tiers in Robotics and AI, 4.

Windridge, D., Felsberg, M., and Shaukat, A. (2012). A

Framework for Hierarchical Perception-Action Lear-

ning Utilizing Fuzzy Reasoning. IEEE transactions

on systems man and cybernetics Part B Cybernetics a

publication of the IEEE Systems Man and Cybernetics

Society, 43(1):155–169.

Wolpert, D. M., Diedrichsen, J., and Flanagan, J. R. (2011).

Principles of sensorimotor learning. Nature reviews.

Neuroscience, 12(12):739–51.

Wolpert, D. M. and Kawato, M. (1998). Multiple paired

forward and inverse models for motor control. Neural

Networks, 11(7-8):1317–1329.

Wolpert, D. M., Miall, R. C., and Kawato, M. (1998). In-

ternal models in the cerebellum. Trends in cognitive

sciences, 2(9):338–47.

Autonomous Vehicle Architecture Inspired by the Neurocognition of Human Driving

513