Why simulated embodied agents?
Embodiment and artificial intelligence
There is an increasing awareness among the scientific community that
genuine intelligence (adaptable, flexible, robust) can emerge only in
a system that is embodied (i.e., has a body through which can interact
with the external environment, using sensors and effectors), and is situated
in an environment it can interact with. The essential implication of embodiment
is the bidirectional, circular interaction between the body of the cognitive
agent and the environment: some of the agent's actions change the state
of the environment, thus changing also the influence of the environment
on the agent (partly perceived through the sensors). This coupling permits
the exploration by the agent of the structure of the environment and the
discovery of structural invariants, through a process which depends on
the sensorimotor capabilities of the agent and its goal. The agent can
thus develop its own conceptualization of the environment, through self-organization
and learning. The grounding of concepts on the sensorimotor interaction
with the environment eliminates the problems of classical AI (lack of
robustness; the lack of access to the semantic content of designer-provided
symbols or categories; the confusion between the agent's perspective and
the observer's perspective).
Simulated embodied agents
While embodiment generally implies a real physical body, like those of
animals and robots, several studies have argued that the importance of
embodiment is not necessarily given by materiality, but by its special
dynamic relation with the environment. This relation can also emerge in
environments other than the material world, such as computational ones.
The environment can be a simulated physical environment, or a genuinely
computational one, such as the internet or an operating system. Simulated
physical environments may be connected to the sensors and effectors of
real physical agents, as in virtual reality, or may also simulate the
body of the agent.
Embodied artificial intelligence research may thus be pursued using either
physical robots or artificial simulated agents (animats). There are several
advantages of using animats. It is much simpler to modify the body of
a simulated agent than to modify a preexisting robot: it may require changing
a few lines of code, versus many hours of engineering work. A simulated
agent may be much cheaper to code, in comparison with the cost of a real
robot. In simulation, one does not have to worry about charging the batteries.
Common real robots have an autonomy of just several hours, when running
on batteries. Simulated robots do not wear off, thus imposing recurrent
costs on the experiment, neither break, which may result in unwanted interruptions
of the experiments. In general, since the hardware considerations may
be omitted, there is more time to focus on the conceptual issues.
Simulation of some simple environments, like in navigation experiments,
may also be faster than real time. This makes simulation preferable for
experiments where the cognitive system of the agent is generated with
evolutive methods, where the behavior of generations of agents in the
environment has to be tracked for long periods of time. Evolutive methods
may also require the repositioning of the agent in the environment, when
starting a new training epoch, which may need to be done manually for
robots, but can be done automatically for animats.
There are also disadvantages of simulation. It is hard to simulate the
dynamics of a physical robot and of an environment realistically, especially
if the simulated agents have many degrees of freedom. In the real world,
the dynamics is simply given by the laws of physics. A simulated environment
is always simpler than the real world, with its infinite richness. This
simplification is based on the designer's perspective of what features
of the environment are important and what are negligible. On one hand,
this limits the possible ontologies that the agent may develop. On the
other hand, it may limit the capability of the agent to deal with the
complexity of the real world.
However, if the purpose of the research is not the design of control
systems that should also work in the real world, but rather the study
of theoretical issues (e.g., sensorimotor integration, the self-organization
of a neural system in interaction with an environment, the grounding of
concepts on the sensorimotor interaction, paradigms for the emergence
of representation in embodied neural systems), simulators are a useful
tool. This paper presents a new simulator adapted for this purpose. It
is a very simple (and thus, fast and convenient) simulator that allows
the study of an agent capable of spatial movement and the manipulation
of discrete objects. The simulator may thus be useful for studies of the
emergence of the concept of object from the sensorimotor interaction,
and also for studies involving navigation or spatial cognitive skills.