Section: New Results

Lifelong Autonomy

Adaptation / Learning

Participant : Jean-Baptiste Mouret.

We collaborate on this subject with Jeff Clune (University of Wyoming, USA).

Adaptation to Unforeseen Damage Conditions

Whereas animals can quickly adapt to injuries, current robots cannot “think outside the box” to find a compensatory behaviour when they are damaged: they are limited to their pre-specified self-sensing abilities and can diagnose only anticipated failure modes, an impracticality for complex robots. A promising approach to reducing robot fragility involves having robots learn appropriate behaviours in response to damage, but current techniques are slow even with small, constrained search spaces. We introduced an intelligent trial- and-error algorithm that allows robots to adapt to damage in less than two minutes in large search spaces without requiring self-diagnosis or pre-specified contingency plans [11] . Before the robot is deployed, it uses a novel technique (based on evolutionary algorithms) to create a detailed map of the space of high-performing behaviours. This map represents the robot's prior knowledge about what behaviours it can perform and their value. When the robot is damaged, it uses this prior knowledge to guide a trial-and-error learning algorithm (based on Bayesian optimization) that conducts intelligent experiments to rapidly discover a behaviour that compensates for the damage. Experiments reveal successful adaptations for a legged robot injured in five different ways, including damaged, broken, and missing legs, and for a robotic arm with joints broken in 14 different ways. This new algorithm will enable more robust, effective, autonomous robots, and may shed light on the principles that animals use to adapt to injury.

This work was the cover of Nature on the 28th of May, 2015 (see the “highlights” section).

Robotics Perception

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

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

Audio Source Localization

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. However, the robot movements are typically fixed or they obey heuristic strategies, such as turning the head and moving towards the source, which may be suboptimal. This work proposes an approach to control the robot movements so as to locate the source as quickly as possible [17] . We represent the belief about the source position by a discrete grid and we introduce a dynamic programming algorithm to find the optimal robot motion minimizing the entropy of the grid. We report initial results in a real environment.

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

State Estimation for Autonomous Surface Vessels

Autonomous Surface Vessels (ASVs) are increasingly proposed as tools to automatize environmental data collection, bathymetric mapping and shoreline monitoring. For many applications it can be assumed that the boat operates on a 2D plane. However, with the involvement of exteroceptive sensors like cameras or laser rangefinders, knowing the 3D pose of the boat becomes critical. We formulated three different algorithms based on 3D extended Kalman filter (EKF) state estimation for ASVs localization [12] . We compared them using field testing results with ground truth measurements, and demonstrated that the best performance is achieved with a model-based solution in combination with a complementary filter for attitude estimation. Furthermore, we presented a parameter identification methodology and showed that it also yielded accurate results when used with inexpensive sensors. Finally, we presented a long-term series (i.e., over a full year) of shoreline monitoring data sets and discussed the need for map maintenance routines based on a variant of the Iterative Closest Point (ICP) algorithm.

Geometric Registration

We proposed a review of geometric registration in robotics [16] . Registration algorithms associate sets of data into a common coordinate system. They have been used extensively in object reconstruction, inspection, medical application, and localization of mobile robotics. We focus on mobile robotics applications in which point clouds are to be registered. While the underlying principle of those algorithms is simple, many variations have been proposed for many different applications. In this work, we gave a historical perspective of the registration problem and showed that the plethora of solutions can be organized and differentiated according to a few elements. Accordingly, we presented a formalization of geometric registration and cast algorithms proposed in the literature into this framework. Finally, we reviewed a few applications of this framework in mobile robotics that cover different kinds of platforms, environments, and tasks. These examples allowed us to study the specific requirements of each use case and the necessary configuration choices leading to the registration implementation. Ultimately, the objective of this work is to provide guidelines for the choice of geometric registration configuration.

Robust Dense Visual Odometry for RGB-D Cameras in a Dynamic Environment

Visual odometry is a fundamental challenge in robotics and computer vision. The aim of our work is to estimate RGB-D camera motion (onboard a mobile robot) from RGB-D images in a dynamic scene with people moving in the scene. Most of the existing methods have a poor localization performance in such case, which makes them inapplicable in real world conditions. This year, we have proposed a new dense visual odometry method [27] that uses random sampling consensus (RANSAC) to cope with dynamic scenes. We show the efficiency and robustness of the proposed method on a large set of experiments in challenging situations and from publicly available benchmark datasets. Additionally, we compare our approach to another state-of-art method based on M-estimator that is used to deal with dynamic scenes. Our method gives similar results on benchmark sequences and better results on our own dataset.

Distributed Sensing and Acting

Participants : Mihai Andries, Amine Boumaza, François Charpillet, Iñaki Fernández Pérez, Nassim Kaldé.

We collaborate on this subject with Olivier Simonin from the Chroma team (Inria Grenoble - Rhône Alpes).

Localisation of Humans, Objects and Robots Interacting on Load-Sensing Floors

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 Hikob company (http://www.hikob.com/ ) and Inria SED (service d'expérimentation et de développement). 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 intelligent apartment experimental platform. This load-sensing floor includes as well a LED lighting system which sits flush with the floor surface. This provides people with a new way to interact with their environment at home. This year, we have focused on localisation, tracking and recognition of humans, objects and robots interacting on load-sensing floors [9] . Inspired by computer vision, the proposed technique processes the floor pressure-image by segmenting the blobs containing objects, tracking them, and recognizing their contents through a mix of inference and combinatorial search. The result lists the probabilities of assignments of known objects to observed blobs. The concept was successfully evaluated in daily life activity scenarii, involving multi-object tracking and recognition on low resolution sensors, crossing of user trajectories, and weight ambiguity.

Online Distributed Learning for a Swarm of Robots

We propose a novel innovation marking method [22] for neuro-evolution of augmenting topologies in embodied evolutionary robotics. This method does not rely on a centralized clock, which makes it well suited for the decentralized nature of embodied evolution where no central evolutionary process governs the adaptation of a team of robots exchanging messages locally. This method is inspired from event dating algorithms, based on logical clocks, that are used in distributed systems, where clock synchronization is not possible. We compare our method to odNEAT, an algorithm in which agents use local time clocks as innovation numbers, on two multi-robot learning tasks: navigation and item collection. Our experiments showed that the proposed method performs as well as odNEAT, with the added benefit that it does not rely on synchronization of clocks and is not affected by time drifts.

The effect of selection pressure on evolution in centralized evolutionary algorithms (EA’s) is relatively well understood. Selection pressure pushes evolution toward better performing individuals. However, distributed EA’s in an Evolutionary Robotics (ER) context differ in that the population is distributed across the agents, and a global vision of all the individuals is not available. In this work, we analyze the influence of selection pressure in such a distributed context. We propose a version of mEDEA [22] that adds a selection pressure, and evaluate its effect on two multi-robot tasks: navigation and obstacle avoidance, and collective foraging. Experiments show that even small intensities of selection pressure lead to good performances, and that performance increases with selection pressure. This is opposed to the lower selection pressure that is usually preferred in centralized approaches to avoid stagnating in local optima.

Online Distributed Exploration of an Unknown Environment by a Swarm of Robots

This year, we have proposed a new taboo-list approach [18] for multi-robot exploration of unknown structured environments, in which robots are implicitly guided in their navigation on a globally shared map. Robots have a local view of their environment, inside which they navigate in a asynchronous manner. When the exploration is complete, robots gather at a rendezvous point. The novelty consists in using a distributed exploration algorithm which is not guided by frontiers to perform this task. Using the Brick and Mortar Improved ant-algorithm as a base, we add robot-perspective vision, variable vision range, and an optimization which prevents agents from going to the rendezvous point before exploration is complete. The algorithm was evaluated in simulation on a set of standard maps.

Another work [14] carried out within the PhD of Nassim Kaldé concerns exploration in populated environments. The difficulty here is that pedestrian flows can severely impact performances. However, humans have adaptive skills for taking advantage of these flows while moving. Therefore, in order to exploit these human abilities, we propose a novel exploration strategy that explicitly allows for human-robot interactions. Our model for exploration in populated environments combines the classical frontier-based strategy with our interactive approach. We implement interactions where robots can locally choose a human guide to follow and define a parametric heuristic to balance interaction and frontier assignments. Finally, we evaluate to which extent human presence impacts our exploration model in terms of coverage ratio, travelled distance and elapsed time to completion.