DA LIO et al.: ARTIFICIAL CO-DRIVERS AS A UNIVERSAL ENABLING TECHNOLOGY 247
reduce speed in curves to maintain the accuracy in lateral
position).
III. C
O-DRIVER OF THE INTERACTIVE PRO JE CT
A. Theoretical Foundations and Design Guidelines
The main goal is to design an agent that is capable of
enacting the “like me” framework, which means that it must
have sensory-motor strategies similar to that of a human and
that it must be capable of using them to mirror human behavior
for inference of intentions and human–machine interaction.
Designing human motor strategies is a relatively easy step:
One may take inspiration from human optimality motor prin-
ciples. For example, we already used the minimum jerk/time
tradeoff and the acceleration willingness envelope to produce
human “reference maneuvers” for advanced driver assistance
systems ( ADAS) [63], [92], [93].
Implementing inference of intentions by mirroring, and
human-peer interactions, is the second less assured step. Two
notable examples (MOSAIC and HAMMER) have been men-
tioned for the general robotics application domain. In the
driving domain, DIPLECS demonstrated learning of the ECOM
structure from human-driving expert-annotated training sets,
and classification of human driver states. However, no co-
driver, in the sense of the definition given in Section I, has been
demonstrated as yet.
The main research question and contribution of this paper is
thus producing a co-driver example implementation to demon-
strate the effectiveness of the simulation/mirroring mechanism
and the following interactions and to focus on the important
potential application impacts that follow from these.
The co-driver has been developed by a combination of direct
synthesis (OC) at the motor primitive level, as well as manual
tuning at higher behavioral levels (the latter being carried
out after inspection of salient situations whenever the two
agents happen to disagree). The final system is thus the cu-
mulation of having compared correct human behaviors (while
discarding incorrect human behaviors) with the developing
agent within many situations encountered during months of
development.
However, in Section VI, we describe how the same archi-
tecture can be potentially employed in the future to implement
“learning by simulation,” namely, optimizing higher-level be-
haviors with simulated interactions via the forward emulators,
to let the system build knowledge automatically (instead of
manually), particularly to accommodate rare events and to
continuously improve its reliability.
Note that, while the main purpose of this system is “un-
derstanding” human goals for preventive safety (see below),
emergency handling and efficient vehicle control intervention
may be added by means of new behaviors (no longer necessarily
human-like) in future versions.
B. Example Implementation
“InteractIVe” is the current flagship project of the European
Commission in the intelligent vehicle domain [99]. It tackles
Fig. 4. Architecture of the co-driver for the CRF implementation of the
Continuous Support function.
vehicle safety in a systematic way by means of three different
subprojects focusing on different time scales: from early holis-
tic preventive safety, to automatic collision avoidance, and to
collision mitigation.
Preventive safety deals with normal driving and with pre-
venting dangerous situations. For this, a “continuous-support
function” has been conceived, which monitors driving and acts
whenever necessary. This functionality integrates, in a unique
human–machine interaction, several distinct forms of the driver
assistance system.
To implement the Continuous Support function, the co-driver
metaphor was adopted. It has been implemented within four
demonstrators of differing kinds. The following describes the
Centro Ricerche Fiat (CRF) implementation, which is closest
to the premises in Section II.
C. Co-Driver Architecture (CRF Implementation)
Fig. 4 shows the adopted architecture. The agent’s “body” is
the car, the agent’s “environment” is the road and its users, and
the “motor output” is the longitudinal and lateral control.
This architecture may be seen to resemble that in Fig. 3,
albeit extended in that the perception–action link is here ex-
plicitly expanded into a subsumptive hierarchy of PA loops.
As indicated in Section II, the input/output structure of layers
within the subsumption hierarchy is characterized by (progres-
sive generalizations of) perceptions and actions. The actual
implementation is built up from functions with input and output
characteristics described in the following section.
By comparison to Fig. 2(b), this architecture is enriched
with forward/inverse models, which make it possible to operate
offline for any purposes requiring “extended deliberation,” e.g.,
for human intention recognition.