Adquisición de conceptos espaciales en un agente autónomo artificial a través de simulaciones internas

In mobile robotics, using artificial vision, one is capable of estimating distance to objects in the environment by means of a camera model. This estimation consists in the calculation of a numerical value based on a number of camera parameters and geometric relations. However, this value needs to be interpreted by the designer or some other external agent. In the work presented here we show how an autonomous agent is capable of acquiring basic spatial concepts about the environment it interacts with. The acquired concepts do not need an external interpretation. In particular, the research focuses on two concepts, the first is distance to, and refers to the distance, in terms of motor commands, the agent must travel before it collides with an obstacle. The second concept is passability , the use of which allows the agent to choose between two passages where only one of them is wide enough for the agent to pass through. These concepts do not rely on a mechanism performing geometrical calculations but on a sensorimotor simulation process coded as an internal model. To achieve this we make use of theories coming from cognitive science, namely we propose the use of forward models. These models, provide the agent with the sensorimotor predictions of its own actions by means of representations constructed on multimodal associations. In this particular case, the system makes use of the visual, tactile and motor modalities, coming from a stereo camera, an array of sonars and the agent actuators, respectively. The work presented is set in the cognitive robotics framework, a research area that holds the nature of human behaviour to be too complex to be implemented as pre-programmed and hardwired by a designer. Instead, the aim is to have agents follow a developmental line similar to that of human beings, learning through interaction with their environment and exploiting the knowledge coming from this interaction. The results obtained in our experiments provide solid ground in the search for more complex agents capable of autonomous interaction, obtaining knowledge grounded on their own sensorimotor capabilities.