Synthesis of robot's behaviors from few examples

Abstract

This paper addresses the problem of the acquisition of robot's behaviors for real environments. It insists on the interest of learning behaviors during robot's interaction with the environment under the control of a human tutor. The paper presents a learning model conceived to synthesize behaviors from a set of few examples and relying on a distributed representation of perception/action relations. The model is experienced on a real robot to learn a slalom task without giving any a priori information about the task or any element of the environment. The model exhibits properties that are well adapted to the interactive learning of concrete behaviors.

Using echo state networks for robot navigation behavior acquisition

“…In order to evaluate the ESN controller approach, all experiments are also performed with two reference algorithms: a naive non-recurrent perceptron with hidden layer and MPL [9], a learning by demonstration algorithm from the literature that is known to be efficient in the context of a mobile robotics with video camera and poor sensor quality. MPL couples a simple yet very efficient image stochastic feature selection algorithm through random sampling of just a few features in the images with a simple correlation-based learning algorithm.…”

“…Two main areas of research can be distinguished : 1) "programming by demonstration" : a human supervisor takes control over the robot (e.g. with a joystick), which is usually considered in the context of mobile robot [13], [9], [12]; 2) "programming by imitation" : a human teacher shows the correct behavior (e.g. using a camera video, laser sensors, etc.…”

“…f (s t ,a t ,s t−1 ,a t−1 ,...,s 0 ,a 0 ) → s t+1 rather than f (s t ,a t ) → s t+1 ). Current and past approaches include recurrent neural networks [3], various flavours of markov models [4] and other approaches based on standard regression setup [9].…”

ISATEM: An Integration of Socioeconomic and Spatial Models for Mineral Resources Exploitation

“…Scope of the ADK approach. The ADK approach is appropriate for applications modelled at a very detailed scale, such as robot foraging problems [5], [11]. In the domain of mineral resources exploitation, it is also appropriate for applications such as the exploitation of quartz at the scale of Rantabe in Madagascar [2] or artisanal and small-scale gold mining at the Alga site in Burkina Faso [3] (let Asgma).…”

“…Designing a SISM involves the selection of the appropriate modelling approach, depending on the application territory scale. A multi-agent system (hereafter MAS) [20] is often used, especially if the application territory is defined at a micro scale [5] [11]. At a larger scale, other approaches, such as dynamic systems [13], may be more appropriate even if a MAS is still acceptable.…”