Section: New Results

Lifelong Autonomy

PsyPhINe: Cogito Ergo Es

Participant : Amine Boumaza.

PsyPhINe is an interdisciplinary and exploratory project (see 8.1.1) between philosophers, psychologists and computer scientists. The goal of the project is related to cognition and behavior. Cognition is a set of processes that are difficult to unite in a general definition. The project aims to explore the idea of assignments of intelligence or intentionality, assuming that our intersubjectivity and our natural tendency to anthropomorphize play a central role: we project onto others parts of our own cognition. To test these hypotheses, our aim is to design a “non-verbal” Turing Test, which satisfies the definitions of our various fields (psychology, philosophy, neuroscience and computer science) using a robotic prototype. Some of the questions that we aim to answer are: is it possible to give the illusion of cognition and/or intelligence through such a technical device? How elaborate must be the control algorithms or “behaviors” of such a device so as to fool test subjects? How many degrees of freedom must it have?

Last year, an experimental robotic device was designed and built, and an experimental campaign with human subject was conducted. The experiments consisted in recording the interactions of the subjects with the robot when realizing a task. The results of the experiments are under analysis and will partly be presented at the second edition of the PsyPhINe workshop organized by the group, gathering top researchers from philosophy, anthropology, psychology and computer science to discuss and exchange on our methodology (see 9.1.1.1).

Localisation of robots on load-sensing floor

Participants : François Charpillet, Francis Colas, Vincent Thomas.

The use of floor-sensors in ambient intelligence contexts began in the late 1990’s. We designed such a sensing floor in Nancy in collaboration with the Hikob company (http://www.hikob.com) and Inria SED. This is a load-sensing floor which is composed of square tiles, each equipped with two ARM processors (Cortex m3 and a8), 4 load cells, and a wired connection to the four neighboring cells. Ninety tiles cover the floor of our experimental platform (HIS).

This year, with Alexis Grall (master student from Enseirb-Matmeca), we have focused on identifying localisation and tracking scenarios involving several robots and on collecting data corresponding to instantiation of these scenarios. These data originated from the sensing tiles but also from Qualisys motion capture system in order to have information about ground-truth. We have also focused on basic algorithms (for instance, Kalman filter) to tackle the issue of tracking targets, but we plan to investigate more elaborate strategies for dealing with sensor discontinuity (for example, when the robot leaves or enters a tile) and multi-traget tracking (Joint Probability Data Association Filter algorithm [52]).

With Mohammad Rami Koujan, we also started to apply deep-learning techniques on those sequential data in order to compare model-based and model-free approaches. This work included some long-term data collection with a randomized behavior in order to have enough training data.

Active sensing and multi-camera tracking

Participants : Olivier Buffet, François Charpillet, Vincent Thomas.

The problem of active sensing is of paramount interest for building self awareness in robotic systems. It consists of a system to make decisions in order to gather information (measured through the entropy of the probability distribution over unknown variables) in an optimal way.

This problem we are focusing on consists of following the trajectories of persons with the help of several controllable cameras in the smart environment. This is a difficult problem since the set of cameras cannot simultaneously cover the whole environment, some persons can be hidden by obstacles or by other persons, and the behavior of each person is governed by internal variables which can only be inferred (such as his motivation or his hunger).

The approach we are working on is based on probabilistic decision processes in partial observability (POMDP - Partially Observable Markov Decision Processes) and particle filters. In the past, we have proposed an original formalism $rho$-POMDP and new algorithms for representing and solving active sensing problems [38] by tracking several persons with fixed camera based on particle filters and Simultaneous Tracking and Activity Recognition approach [45].

This year, we have focused on investigating the issue of solving the active sensing problem with controllable cameras. Approaches based on Monte-Carlo Tree Search algorithms (MCTS) like POMCP [54] are currently investigated for adressing the combinatorial explosion of the state space to consider (which is the space of probability distributions over all the possible states of the system).

Audio Source Localization

Participants : François Charpillet, Francis Colas, Van Quan Nguyen.

We collaborate on this subject with Emmanuel Vincent from the Multispeech team (Inria Nancy - Grand Est).

We considered, here, the task of audio source localization using a microphone array on a mobile robot. Active localization algorithms have been proposed in the literature that can estimate the 3D position of a source by fusing the measurements taken for different poses of the robot. A typical implicit assumption in the literature is that the sound source is active, but a lot of real sound sources are actually intermittent. Systems of activity detection exist but cannot reach perfect accuracy. In this work, we propose a new mixture Kalman filter that explicitly includes the discrete activity of the source in the estimated state vector, alongside the continuous states such as the position of the robot or the sound source. We take into account the imperfection of activity detection systems in order to show that we have better accuracy than the state of the art [26].

This work is led through the PhD Thesis of Van Quan Nguyen under the supervision of Emmanuel Vincent and Francis Colas.

Learning for damage recovery

Participants : Jean-Baptiste Mouret, Konstantinos Chatzilygeroudis, Vassilis Vassiliades, Dorian Goepp.

In 2015, we introduced a novel algorithm that allows robots to learn by trial-and-error when they are damaged [42]. In 2016, we extended this algorithm to make it easier to deploy it in real-life situations and real systems:

• We added “safety constraints” so that the learning algorithm both maximizes the post-damage performance and minimizes the probability of breaking the robot during the learning process; we demonstrated this extension with a simulation of the iCub robot, which is a fragile and expensive robot (around 250,000 euros) for which we would like to use our learning algorithms [33].

• We proposed a novel algorithm that does not require to reset the robot to its starting position between each trial [40], which allows the damaged robots to “learn while doing”. We demonstrated this algorithm on our 6-legged walking robot.

• We extended the MAP-Elites algorithm, that is, the evolutionary algorithm that we use to generate prior probability distributions for our online learning algorithm, to scale-up to high-dimensional search spaces [59]. The algorithm is based on central Voronoi tesselations (CVT). In addition, we investigated the influence of the encoding (representation of the controller) on the performance of MAP-Elites [30].

Interactions with biology

Participant : Jean-Baptiste Mouret.

We continued our on-going collaborations with biologists.

• The Evolutionary Origins of Hierarchy. Hierarchical organization—the recursive composition of sub-modules—is ubiquitous in biological networks, including neural, metabolic, ecological, and genetic regulatory networks, and in man-made systems such as large organizations and the Internet. In this contribution, we showed that the pressure to minimize the connection costs in network can explain the evolution of hierarchical and modular biological networks [21]; this result extends our previous work on the evolutionary origins of modularity in biological networks [41]. (Collaboration with Jeff Clune, University of Wyoming, USA).

• Animal-robot interaction. We worked with a team based in Rennes to perform preliminary experiments about animal-robot attachment (here with a gallinaceous bird) [16].

Learning for whole-body motions

Participants : Serena Ivaldi, Valerio Modugno.

Within the European project CoDyCo, we studied how to combine learning, dynamics, and control for redundant robots. In [25], we proposed a framework to automatically optimize the evolution in time of soft task priorities for multi-task controllers. The motivation of the work was to propose a way to automatate the manual optimization procedure of task priorities and weights, that is classically done by control experts and is time consuming. In [24], we improved the framework by using constrained stochastic optimization algorithms to optimize the task priorities while ensuring that the system constraints (robot and problem setting) are never violated. We showed the results on our robot iCub. Our master student Ugo Chervet contributed to the simulations of this paper.