Section: New Results

Autonomous And Social Perceptual Learning

Unsupervised and online non-stationary obstacle discovery and modelling using a laser range finder

Participants : Guillaume Duceux, David Filliat [correspondant] .

Recognizing objects is an important capability for assistance robots, but most methods rely on vision and a heavy training procedures to be able to recognize some objects. Using laser range finders has shown its efficiency to perform mapping and navigation for mobile robots. However, most of existing methods assume a mostly static world and filter away dynamic aspects while those dynamic aspects are often caused by non-stationary objects which may be important for the robot task. We propose an approach that makes it possible to detect, learn and recognize these objects through a multi-view model, using only a planar laser range finder. We show using a supervised approach that despite the limited information provided by the sensor, it is possible to recognize efficiently up to 22 different object, with a low computing cost while taking advantage of the large field of view of the sensor. We also propose an online, incremental and unsupervised approach that make it possible to continuously discover and learn all kind of dynamic elements encountered by the robot including people and objects. These results have been published at the IROS conference [40] .

Task oriented representations by discriminative modulation of a generative learning method

Participants : Mathieu Lefort, Alexander Gepperth [correspondant] .

PROPRE (which stands for PROjection - PREdiction) is a generic and modular unsupervised neural learning paradigm that extracts meaningful concepts of multiple data flows based on predictability across stimuli. It consists on the combination of three modules. First, a topological projection of each data flow on a self-organizing map. Second, a decentralized prediction of each projection activity from each other map activities. Third, a predictability measure that quantifies the prediction error. This measure is used to modulate the projection learning so that to favor the mapping of predictable stimuli across data flows. This model was applied to the visual supervised classification of the pedestrian orientation. The modulation of the visual representation learning by the predictability measure (quantifying the ability to detect the orientation of the pedestrian) improves significantly classification performances of the system independently of the predictability measure used [55] . Moreover, PROPRE provides a combination of interesting functional properties, such as online and incremental learning [56] .

Learning of multimodal representations based on the self-evaluation of their predictability power

Participants : Mathieu Lefort, Thomas Kopinski, Alexander Gepperth [correspondant] .

PROPRE paradigm (see section  6.4.2 ) was also applied to the classification of gestures caught from two time-of-flight (ToF) cameras. In this context, the predictability measure acts as a self-evaluation module that biases the learned representations towards stimuli correlated across modalities, i.e. related to the ability of one camera to predict the other one. We show in [57] that this unsupervised multimodal representations learning improves the gesture recognition performance, compared to isolated camera representations learning, even not as much as supervised one.

Resource-efficient online learning of classification and regression tasks

Participants : Mathieu Lefort, Thomas Kopinski, Thomas Hecht, Alexander Gepperth [correspondant] .

This activity investigates the coupling of generative and discriminative learning (SOM and regression) to achieve incremental learning that stays resource-efficient when the number of input and output dimensions is high. On the one hand, we apply this technique to sensory classification problems where input dimensionalities can exceed 10000 in the presence of multiple categories. On the other hand, we target the learning of forward and inverse regression models for robotics, possibly combining proproceptive with sensory information which again leads to high data dimensionality. A special kind of regression task we consider in this context is optimal integration of sensory information, where the most likely underlying value must be inferred from several noisy sensor readings. In contrast to popular approaches like XCF or LWPR, our approach achieves efficiency by avoiding a precise partitioning of the input space, relying on a dimensionality-reduced topological projection of the input space instead. While this achives slightly inferior results on standard benchmarks, we can treat high-dimensional incremental learning problems that are inaccessible to other algorithms, and especially to LWPR. This activity has resulted in two submissions to ESANN 2015 and one to IEEE Transactions on Autonomous Mental Development.

Indoor semantic mapping on a mobile robot

Participants : Louis-Charles Caron [correspondant] , Alexander Gepperth, David Filliat.

Semantic mapping is the act of storing high-level information in a persistent map of the environment. The semantic information considered here is the identity of objects encountered by a mobile robot in an indoor environment [35] . The robot runs a SLAM algorithm and builds a map using a laser range finder. The semantic information is collected by analysing the point cloud provided by an RGB-D camera mounted on the robot. The choice of features used to describe the objects, the type of fusion and the recognition algorithm influence the overall capacity of the algorithm. Shape features perform very well, but are blind to changes in color. The fusion of different types of features can reduce the recognition rates on some objects but increases the overall figure. This increase is more significant as the number of objects to recognitze gets larger [36] . After running the object recognition algorithm, the identity of the objects is stored alongside the map. The stored information influences future recognition attempts on objects that were already seen by the robot to improve the recognition process. A 3-d map along with a snapshot and the identity of each object seen is displayed to a user.