This research is done in collaboration with Renault as part of a CIFRE PhD fellowship that started in June 2009. It was done also in collaboration with Prof. S. Thrun's Driving Group at the Stanford Artificial Intelligence Lab. The recent advances in V2X communication technologies allow the sharing of information among vehicles (V2V) and between vehicles and road infrastructure (V2I). The shared information can be used locally by each entity to extend its awareness horizon and build a better representation of the surrounding environment. Based on this principle Advanced Driver Assistance Systems (ADAS) can be implemented to enhance driving comfort and safety. The focus of this PhD is on cooperative safety applications at intersections (e.g. violation warning, collision warning, intersection assistant).

Intersections are among the most complex environments encountered in road networks. The large number of geometrical configurations, signalization, traffic rules, and vehicle interactions result in many different possible scenarios. Situation assessment is therefore both difficult and fundamental for the safe crossing of intersections. Difficult since the large number of potential scenarios makes it hard to interpret what is happening. Fundamental as any misinterpretation could lead to hazardous situations resulting in accidents. The key issue (and challenge) for situation understanding at intersections is to be able to infer vehicle behavior from incomplete models and uncertain data. In our case the objective is to build a probabilistic model that can estimate the behavior of vehicles at an intersection using:

Therefore for this project on Cooperative Vehicle Applications the first step is to work on situation assessment (e.g. determine if a vehicle is going to stop at an intersection?) and the second one is to do risk assessment (e.g. determine the risk of collision between two vehicles?).

In 2010 we developed a system that estimates a vehicle's intended destination at an intersection based on its current state and on contextual information extracted from the digital map (see illustration in Figure ). The idea is to use the information on the geometry of the road network and on the connectivity between lanes to build a statistical model of the relationship between the position and turn signal of a vehicle and its intended destination. The model is a Bayesian Network where the values in the conditional probability tables are set using the information extracted from the map. The system is generic to any intersection layout and is capable of handling uncertain input information. So far we conducted experimental evaluation using a set of 42 recorded trajectories of vehicles at two different intersections. The performances were evaluated by measuring how early on average the system made a correct prediction about the destination of a vehicle in five different scenarios including complex situations where the driver’s behavior was inconsistent (e.g. the driver put on the left turn signal and turned right). A qualitative evaluation showed that the system handles the different scenarios in a coherent manner even when the driver’s behavior is inconsistent. We were also able to demonstrate the advantages of our approach compared with approaches that do not integrate contextual information and do not handle input information uncertainties. This work, in collaboration with the Stanford Artificial Intelligence Lab, resulted in the implementation of the proposed system as a module in the Driving Group's autonomous driving framework. An article describing this work was submitted in November to the IEEE CIVTS'11 symposium .

Two conference articles describing the results and findings after the participation of Renault to the European project Safespot (“Cooperative vehicles and road infrastructure for road safety”) were published in 2010 , . The four Use Cases implemented on the Renault demonstrators were a) Warning - Emergency vehicle approaching an intersection, b) Warning - Emergency vehicle coming from behind, c) Warning - Accident at intersection, d) Warning - Crash predicted at intersection. They are illustrated in Figure .

An important goal of the ArosDyn project is to develop a system which integrates our recently developped techniques and provides a real-time collision risk estimation in a dynamic environment. To achieve this goal, we designed and implemented the ArosdynTestSuite software. The main features of this software are:

The architecture of this software is shown in Fig. .

In this example, we demonstrate a typical sensor fusion application. We retrieve the raw data from the Hugr middleware and store them in individual sensor objects. Then, by using this framework, we integrate the IBEO Bayesian Occupancy Filter (BOF) sensor model, the stereo sensor processor model, the stereo BOF sensor model and the BOF model together. Finally, different aspects of the computational results are visualized in several viewers. At the same time, all the parameters used by the algorithms can be tuned online.

Several windows of this application are shown in Fig. . Here we demonstrate the main window, the 2D viewer of the stereo camera and the lidar, the disparity map of the stereo vision and the compounded BOF grid which is the result of the sensor fusion.

Another important property of this software is a large part of the computation task executed on GPU. As the processing of stereo image and the computaion in the BOF can be highly parallelized, we run these tasks on the GPU to improve the time performance, as shown in Fig. . In this way, the software can work in real-time.

The GPU calculation is based on CUDA library and is carried out in an independent thread. The schematic graph of the GPU computaional thread is shown in Fig. .

Furthermore, thanks to the deliberated design of the software, we can easily add new models to it and let them work together. The fast detection and tracking algorithm (FCTA) and the Gaussian process based collision assessment algorithm will be added into this framework.

The ArosDyn project (an INRIA Technological Developpement Action) aims to develop embedded software for robust analysis of dynamic scenes in urban traffic environments, in order to estimate and predict collision risks during car driving. The on-board telemetric sensors (lidars) and visual sensors (stereo camera) are used to monitor the environment around the car. The algorithms make use of Bayesian fusion of heterogenous sensor data. The key objective is to process sensor data for detection and tracking of multiple moving objects for estimating and predicting collision risks in real time, in order to help avoid potentially dangerous situations  , . The local environment is represented by a grid, and the fusion of sensor data is accomplished by means of the Bayesian Occupancy Filter (BOF)  , , that provides to assign probabilities of cell occupancy and cell velocity for each cell in the grid. The collision risks are considered as stochastic variables. Hidden Markov Model (HMM) and Gaussian Process (GP) are used to estimate and predict collision risks and the likely behaviours of multiple dynamic agents in road scenes for a short period ahead.

In many computer vision related applications it is necessary to distinguish between the background of an image and the objects that are contained in it. This is a difficult problem because of the double constraint on the available time and the computational cost of robust object extraction algorithms.

For the specific problem of finding moving objects from static cameras, the traditional segmentation approach is to separate pixels into two classes: background and foreground. This is called Background Subtraction  and constitutes an active research domain (see ). Having the classified pixels, the next processing step consists of merging foreground pixels to form bigger groups corresponding to candidate objects, this process is known as object extraction.

In previous work, we focused on the background/foreground classifier exploring the new capability of the extraction algorithm. We implemented a statistical classifier that models the background pixels as Gaussians and we performed the foreground classification based on the Mahalanobis distance between the intensity of the input pixel and its correspondent background model.

In 2009, we adopt the Mixture of Gaussians (MoG) to model the background, We compared the results on combining our clustering algorithm with 4 different background/foreground classification techniques. In 2010 we did more experiments and published in IROS conference .

Obstacle detection is a widely explored domain of mobile robotics. It presents a particular interest for the intelligent vehicle community, as it is an essential building block for Advanced Driving Assistance Systems (ADAS). In the LOVe project, the E-Motion team proposes to perform obstacle detection within the occupancy grid framework. For this purpose, the Bayesian Occupancy Filter (BOF) was presented as previous work . Its performances and functionalities were demonstrated in particular with data from laser scanner.

To use other sensors in this framework, it is essential to develop an associated probabilistic sensor model that take into consideration the uncertainty over measurements. In 2009, we proposed such a sensor model for stereo-vision . The originality of the approach relies on the decision to work in the disparity space, instead of the classical metric space. This idea gives two major improvements. First, the errors on measurements are more accurately modeled. In particular, the proposed Gaussian model takes into account that the measurement uncertainty is related to the range of observed object. Second, the use of accumulation methods in the disparity space, such as the u-v-disparity approach , helps to be computationally efficient.

In the LOVe project, our novel method has been applied to real urban dataset for obstacle detection and tracking. Good detection rates were obtained, even in very dynamic and rich environment such as the downtown driving context. It also appeared from the study of the results that our Gaussian sensor model made it easier to track moving objects using the previously proposed Fast Clustering Tracking Algorithm .

In 2010, we improved our sensor model, in order to mimic some features of the sensor models used for range finders. Particularily, we worked on managing visible/occluded areas of the scene. For this purpose, working in the disparity space is still a good way to proceed: the u-disparity plane provides a polar representation in which all the rays of light coming through the camera matrix are parralels. Occlusions and uncertainty are very easy to handle with this representation, thus we propose a probabilistic approach based on this idea . Our approach provides objective advantages over other methods from the state of the art. Moreover, it allows to perform higly parralel computation of the occupancy grid: A.Nègre implemented the approach on GPU using NVIDIA CUDA, reaching very high performances. The complete processing of stereo data can now be done in 6 ms, while more than 150 ms were necessary with the CPU implementation.

In this approach, we only considered the measurements belonging to actual obstacles. Since stereo-vision allows to classify pixels as "road" or "obstacles", the modelization of the scene could be improved by adding the "road" information. We proposed a way to integrate this information in .

The sensor model for stereo-vision is used for dynamic scene analysis in the Arosdyn project . In this project, we intend to perform sensor fusion in the occupancy grid framework, using both lidars and stereo-vision. We conducted some early experiments on this approach of sensor fusion, using real data aquired from our experimental vehicle .

Figure shows an exemple of occupancy grid computed using our new approach. We can observe that most objects are detected (light color), even if partially occluded (e.g. the sign on the right). Information from the road surface is also taken into consideration (dark areas). Moreover, similar to a laser scanner, it appears that regions in front of objects are seen as partially unoccupied, while less informations is available behind obstacles (occupancy probability is closer to 0.5).

This is a new activity which includes and generalizes some of the results obtained for a specific estimation problem (the problem of sensor calibration). During 2009, the problem of sensor self-calibration in mobile robotics by only using a single feature (e.g. a vertical or a horizontal line) was investigated. The main results achieved during that year were published in , , and .

During 2010 a general framework which applies in any estimation problem has been developed. In particular, the investigation regarded problems where the information provided by the sensor data is not sufficient to carry out the state estimation (i.e. the state is not observable). For these systems the concept of continuous symmetry has been introduced. Detecting the continuous symmetries of a given system has a very practical importance. It allows us to detect an observable state whose components are non linear functions of the original non observable state. This theoretical and very general concept has been applied to deal with two distinct fundamental estimation problems in mobile robotics. The former is in the framework of self-calibration and the latter is in the framework of the fusion of the data provided by inertial sensors and vision sensors. For both problems all the observable modes were analytically derived by analyzing the continuous symmetries. Finally, for both problems several closed-form solution have been analytically derived. Some of the results have been published in .

This activity has been carried out in the framework of the European project sFly. In particular, we both considered the case of a single vehicle and the case of a team of cooperating vehicles.

Regarding the case of one vehicle, we introduced a new approach to solve SLAM. This approach combines an Information Filter and a non linear optimizer. The basic idea of the suggested technique is to use the Information Filter when the system non linearities are negligible, and to switch to the use of the non linear optimizer when the non linearities are not negligible. Extensive simulations have been performed in order to evaluate the performance of the proposed approach. In particular, a comparison with the Exactly Sparse Delayed-state Filers (ESDF) technique has been carried out. The main results have been published and are available in .

Regarding the case of a team of cooperating vehicles, we introduced a new approach to the problem of simultaneously localizing micro aerial vehicles (MAV) equipped with inertial sensors able to monitor their motion and with exteroceptive sensors. The method estimates a delayed state containing the trajectories of all the MAVs. The estimation is based on an Extended Information Filter whose implementation is distributed over the team members. The approach contains two novel contributions. The former is a trick which allows exploiting the information contained in the inertial sensor data in a distributed manner. The latter is the use of a projection filter which allows exploiting the information contained in the geometrical constraints which arise as soon as the MAV orientations are characterized by unitary quaternions. The performance of the proposed strategy has been evaluated with synthetic data. In particular, the benefit of the previous two contributions has been pointed out.

The salient features of our algorithm are as follows:

The main results have been published in .

Perceiving or understanding the environment surrounding a vehicle is a very important step in driving assistance systems or autonomous vehicles. The task involves both Simultaneous Localization And Mapping (SLAM) and Detection And Tracking of Moving Objects (DATMO). In this context, we have designed and developed a generic architecture to solve SLAM and DATMO in dynamic outdoor environments.

For the SLAM problem, this architecture uses a maximum likelihood approach to build a consistent local map using occupancy grid and to localize the ego vehicle inside the map. After a consistent local map has been constructed, moving objects can be detected using inconsistencies between observed free space and occupied space in the local grid map .

This year was focused on the integration of vision information and their fusion with laser data to improve the quality of the DATMO process. The Fusion process takes as an input the list of objects provided by both kind of sensors and delivers a fused list of objects where for each object we have information about its pattern and its class. This work has been done in the frame of the european project INTERSAFE2.

Figure  shows the results of the process. In green color, we see the results of the process: all the moving objects present in the field of view of the vision system (green color field of view) are detected by our system.

Navigation in large dynamic spaces has been often adressed using deterministic representations, fast updating and reactive avoidance strategies. However, probabilistic representations are much more informative and their use in mapping and prediction methods improves the quality of obtained results. We propose a new concept to integrate a probabilist collision risk function linking planning and navigation methods with the perception and the prediction of the dynamic environments. Moving obstacles are supposed to move along typical motion patterns represented by Gaussian Processes. The likelihood of the obstacles' future trajectory and the probability of occupation are used to compute the risk of collision. The proposed planning algorithm is a sampling-based partial planner guided by the risk of collision. The perception and prediction information are updated on-line and reused by the planner. The decision takes into account the most recent estimation. In 2008, we integrated our search algorithm with a representation of typical patterns based on Markov graphs and started to use another representation based on Gaussian Processes . In 2009, we proposed more complete results published in IROS 2009 where a robot need to reach goals in the hall of INRIA avoiding pedestrians. Chiara fulgenzi defended her Ph.D on June 2009 . In 2010 the Risk-RRT Algorithm was integrated in the cycabtk simulator. Results show the performance for a robotic wheelchair in a simulated environment among multiple dynamic obstacles (see fig ). These results have been submitted to IEEE Transactions on Robotics and also published as a research report On going works focus on adapting the navigation method to the wheelchair taking into consideration that this wheelchair navigates in a populated environment AND transports a person. Next section will details the main points of these research.

The problem adressed in this work is the autonomous navigation of wheelchairs. If one consider that the navigation system transports a person, the integration of social conventions in the navigation strategy starts to be crucial. In this work, we propose to integrate the notions of personal space and interaction between people. We propose to enrich the knowlegde the robot has, with a representation of the social conventions. The wheelchair must take into consideration interactions to avoid groups of people (even if passing through the group is the “best ”path for a conventionnal planning algorithm), or to join a group with a behavior close to the one of a human. To understand the behaviors of interaction between humans and the management of space we can support us on the works developed in the area of sociology to define some concepts as Personal space, o-spaceand F-formations.

That definition is important because it could be a useful tool for a robot to understand the intentions of the humans. It is well known that these measures are not stricts and that they change depending on age, culture and type of relationship but the categories proposed explain very well reactions like the uncomfortable sense of a stranger invading your intimate zone or the perception of somebody looking social interaction because he is entering to your social zone.

This research has been carried out in the framework of the European project sFly. In recent years it is revealed more and more the importance of using multi-robot systems for security application, otherwise impossible to be performed by a single robot. In particular by employing flying robots, many fundamental tasks are now possible. Some of these very important tasks are: surveillance of dangerous regions, like areas of chemical, biological or nuclear contamination; environmental monitoring ( egair quality, forest fire); aiding police during surveillance missions and so on. For these purposes we try to develop efficient, robust and versatile methodologies for distributed control of multi-MAV systems under environmental and communication constraints such as global path, obstacles, loss of local communication.

The first problem approached is the optimal coverage of an unknown environment, i.e. finding the optimal deployment for the robots' team to monitor a given region. The solution for the 2-D case without obstacles is already known in literature and it can be obtained analytically. It is based on the Voronoi partition generated by the positions of the robots and the Loyd algorithm  . For the non-convex case, i.e. an environment with obstacles, in we firstly proposed a possible strategy based on a combination of the repulsive potential field method and the Voronoi partition. In other words the movement of each robot is generated by the repulsion of the other robots, by the repulsion of the closest obstacle and by the attraction of the center of mass of its Voronoi region. This method allows overcoming the problem of local minima and the robots are better spread in the environment. In this year we have mainly approached the coverage problem by using a new stochastic optimization method. This work is in collaboration with professor Elias Kosmatopoulos and his group, of the Technical University of Crete, partner in the sFly project.

The Kosmatopoulos's group has proposed a new approach for developing efficient and scalable methodologies for a general class of multi-robot passive and active sensing applications  , . This method employs an estimation scheme that switches among linear elements and, as a result, its computational requirements are about the same as those of a linear estimator. The parameters of the switching estimator are calculated off-line using a convex optimization algorithm which is based on optimization and approximation using Sum-of-Squares (SoS) polynomials. Stable and convergent estimator's performance is guaranteed, overcoming the shortcoming of many existing methodologies where there is always the possibility of estimator error divergence.

The main results obtained for the 2-D case by using this method has been published in , . We assume the robots are equipped with global positioning capabilities and visual sensors able to monitor the surrounding environment. The goal is to maximize the area monitored by the team, by identifying the best configuration of the team members. The proposed approach has the following key advantages with respect to previous works:

The previous approach has been then extended for the 3-D case and some simulations using real data, which were collected with the use of a miniature quadrotor helicopter specially designed for the needs of the European project sFLY, have been performed. We developed also a distributed version of this method to solve the same problem. In multi-robot systems, a distributed approach is desirable for several fundamental reasons. The most important are failure of the central station and limited communication capabilities. Furthermore, several simulations show that in our case the distributed algorithm is faster to converge than the centralized one. The results of this work led to the submission of another paper, now under review.

Finally, in we have proposed also a new distributed algorithm for cooperative exploration of an unknown environment. The proposed approach is based on the potential field method. The advantages of using this method are several and well known, but the presence of many local minima does not assure the exploration of the entire environment. Our idea is to preserve these advantages but overcome the problem of local minima by introducing a leader in the team which has a different control law, unaffected by this problem. Furthermore, we consider also the case of several local leaders, dynamically selected on the basis of a hierarchy within the team.

Where to move next?is a key question for an autonomous robotic system. This fundamental issue has been largely addressed in the past forty years. Many motion determination strategies have been proposed. They can broadly be classified into deliberativeversus reactivestrategies: deliberative strategies aim at computing a complete motion all the way to the goal, whereas reactive strategies determine the motion to execute during the next time-step only. Deliberative strategies have to solve a motion planning problem  . They require a model of the environment as complete as possible and their intrinsic complexity is such that it may preclude their application in dynamic environments. Reactive strategies on the other hand can operate on-line using local sensor information: they can be used in any kind of environment whether unknown, changing or dynamic, but convergence towards the goal is difficult to guarantee.

To bridge the gap between deliberative and reactive approaches, a complementary approach has been proposed based upon motion deformation. The principle is simple: a complete motion to the goal is computed first using a priori information. It is then passed on to the robotic system for execution. During the course of the execution, the still-to-be-executed part of the motion is continuously deformed in response to sensor information acquired on-line, thus accounting for the incompleteness and inaccuracies of the a priori world model. Deformation usually results from the application of constraints both external (imposed by the obstacles) and internal (to maintain motion feasibility and connectivity). Provided that the motion connectivity can be maintained, convergence towards the goal is achieved.

The different motion deformation techniques that have been proposed  , , , , all performs path deformation. In other words, what is deformed is a geometric curve, iethe sequence of positions that the robotic system is to take in order to reach its goal. The problem with path deformation techniques is that, by design, they cannot take into account the time dimension of a dynamic environment. However in many cases, it would be more appropriate to try to anticipate the motion of dynamic obstacles by considering their current and future motion, in the aim to avoid collision with them in a close future. To achieve this, it is necessary to depart from the path deformation paradigm and resort to trajectory deformationinstead. A trajectory is essentially a geometric path parametrized by time. It tells us where the robotic system should be but also when and with what velocity. By taking into account a forecast model of the future behavior of obstacles, both spatial and temporaldeformations may take place. They result from the application of external forces ierepulsive forces exerted from the obstacles and internal forcesused to maintain the connectivity of the trajectory. Both path and velocity profiles of the robotic system may thus be altered to adapt the behavior of the robot to its environment.

Our trajectory deformer named Teddyis designed to be one component of an otherwise complete autonomous navigation architecture. A motion planning module is required to provide Teddywith the nominal trajectory to be deformed. Teddyoperates periodically with a given time period. At each cycle, Teddyoutputs a deformed trajectory which is passed to a motion control module that determines the actual commands for the actuators of the robotic system. The work on Teddyculminated in 2010 with the completion of Vivien Delsart's PhD.

This research topic is related to our work on Trajectory Deformation (see Section  ). The main difficulty of trajectory deformation lies in the maintenance of the connectivity of the trajectory for complex dynamic systems since it requires the characterization of their set of reachable states. To solve this problem, we have decided instead to base the internal forces computation upon a steering methodcomputed by a new trajectory generatorcalled Tiji.

Trajectory generation for a given robotic system is the problem of determining a feasible trajectory (that respects the system's dynamics) between an initial and a final state. From the preliminary work of Dubins  to the latest methods used by the Carnegie Mellon University during the Darpa Urban Challenge  , many trajectory generation methods have been proposed like primitive combinations  , , , , two-point boundary value problems  , , , or variational approaches  , , , .

Among all these approaches, it is interesting to note that, in no circumstances, have people tried to compute a trajectory reaching the goal state at a specific time instant. However, several problems like Trajectory Deformation require to fix the final time or at least an interval of time during which a goal state must be reached. We proposed thus a trajectory generation scheme called Tijiwhich integrates the final time constraint. Tijiis geared towards complex dynamic systems subject to differential constraints, such as car-like vehicles, and its efficiency warrants it can be used in real-time. The approach is similar in spirit to that of  : a parametric trajectory representation is assumed in order to reduce the search space. An initial set of parameters is selected yielding a trajectory that does not necessarily reach the goal state. The parameter space is then searched and efficient numerical optimization is used to optimize a cost function involving the distance between the end of the trajectory computed and the (goal state, final time) pair. Should the goal state be unreachable (if the final range of time is ill-chosen), the method returns a trajectory that ends as close as possible to the (goal state, final time) pair. Tijidiffers from previous works first because it takes into account the final time constraint and also because the control definition chosen is such that it ensures that all trajectory constraints are met. The latest results concerning have been reported in  . They are also included in Vivien Delsart's PhD.

Safe navigation is undoubtedly a critical issue that needs to be solved for autonomous mobile robots/vehicles to leave labs environments and be deployed in real world applications. This problem has been studied extensively by the robotics community and some results were illustrated rather brilliantly by the 2007 DARPA Urban Challenge http:// www. darpa. mil/ grandchallenge.. The challenge called for autonomous car-like vehicles to drive 96 kilometers through an urban environment amidst other vehicles. Six autonomous vehicles finished the race thus proving that autonomous urban driving could become a reality. Note however that, despite their strengths, the Urban Challenge vehicles have not yet met the challenge of fully autonomous urban driving (how about handling traffic lights or pedestrians for instance?). Moreover, at least one collision took place between two competitors. It shows that motion safety(the ability for an autonomous robotic system to avoid collision with the objects of its environment) remains an open problem in mobile robotics.

To address this issue, we have explored the novel concept of Inevitable Collision States(ICS) since 2002. An ICS for a given robotic system is a state for which, no matter what the future trajectory followed by the system is, a collision with an object eventually occurs. For obvious safety reasons, a robotic system should never ever end up in an ICS. ICS have already been used in a number of applications: (i) mobile robot subject to sensing constraints, iea limited field of view, and moving in a partially known static environment  , (ii) car-like vehicle moving in a roadway-like environment  , (iii) spaceship moving in an asteroid field  . In all cases, the future motion of the robotic system at hand is computed so as to keep the system away from ICS. To that end, an ICS-Checker is used. As the name suggests, it is an algorithm that determines whether a given state is an ICS or not. Similar to a Collision-Checker that plays a key role in path planning and navigation in static environments, it could be argued that an ICS-Checker is a fundamental tool for motion planning and navigation in dynamic environments. Like its static counterpart, an ICS-Checker must be computationally efficient so that it can meet the real-time constraint imposed by dynamic environments.

This work has yielded first a genericand efficientICS-Checker for planar robotic systems with arbitrary dynamics moving in dynamic environments, and second, an ICS-based collision avoidance scheme, iea decision-making module whose primary task is to keep the robotic system at hand safe from collisions. This collision avoidance scheme, guarantees motion safety with respect to the model of the future which is used. They have been documented in 2010 in Luis Martinez-Gomez's PhD.

Finally we have addressed how the ICS concept can be extended to handle the uncertainty inherent to the future. So far the characterization of the ICS has been based upon deterministic models of the future. In other words, each moving object was assumed to follow a given nominal trajectory (known a priori or predicted). Such deterministic models provide a clear-cut answer to the motion safety issue: a given state is an ICS or not (simple binary answer). However, such models are not well suited to capture the uncertainty that prevails in real world situations, in particular the uncertainty concerning the future behaviour of the moving objects. Our contribution is a probabilistic formulation of the ICS concept. Probabilistic ICS permit the characterization of the motion safety likelihood of a given state, likelihood that can later be used to design safe navigation strategies in real world situations. This is the first step towards the applicability of the ICS framework to real robots operating in uncertain dynamics environments. The first results concerning Probabilistic ICS have been reported in  .

Nowadays, robotic systems are increasingly moving among and interacting with human beings. The presence of people has an impact on the way robots should move: you do not move around a person the way you would around a piece of furniture. In general, when one interacts with a human being (to avoid him, to salute him, to pass him an object, etc.), one obeys a number of unspoken social rules, egmaintaining an appropriate distance. These rules must be identified and integrated in the navigation scheme of robotic systems so that their behaviour resemble that of a person thus facilitating the integration of robots in human-populated environments and their acceptance.

The primary contribution of this novel line of research is a navigation scheme that is anthropomorphic, iethat emulates human behaviors and seeks to adhere to these social rules. Unlike previous works in this area, the focus herein is on dynamic scenarios. The navigation scheme proposed explicitly reasons on the future behavior of the people involved so as to produce better socially acceptable trajectories (not to mention safer trajectories as well). The navigation scheme relies upon a novel cost function called the social costmapthat captures in a unified way the different social rules imposed by the people populating the robot's workspace (see Fig.  ). The first results concerning anthropomorphic navigation will be presented during the 2011 IEEE Int. Conf. on Robotics and Automation.

This work has been done in collaboration with our Start-up Probayes. ProBT is a C++ library for developing efficient Bayesian software . This library has two main components: (i) a friendly Application Program Interface (API) for building Bayesian models and (ii) a high-performance Bayesian inference and learning engine allowing execution of the probability calculus in exact or approximate ways.

The aim of ProBT is to provide a programming tool that facilitates the creation of Bayesian models and their reusability ( http:// www. probayes. com/ ). Its main idea is to use “probability expressions” as basic bricks to build more complex probabilistic models. The numerical evaluation of these expressions is accomplished just-in-time: computation is done when numerical representations of the corresponding target distributions are required. This property allows designing advanced features such as submodel reuse and distributed inference. Therefore, constructing symbolic representations of expressions is a central issue in ProBT.

Since a few years the main development of ProBT has been carried out by Probayes: a spin-off born from the e-motion project. Both, e-motion and Probayes, are part of the European project Bayesian Approach to Cognitive Systems (BACS). The development of ProBT has been done taking into account the goals of the BACS project partners.

The problem addressed in this work is the autonomous replacement of a human player. We worked on two types of games, and so two types of players. From a research point of view, multiplayer games are interesting because they stand for a good in-between of the real world and simulations. The world is finited and simulated (no sensors problems) but we didn't wrote the simulation and the other players are humans (or advanced robots in the case of AI competitions).

The first work was on the control of a virtual avatar by an artificial intelligence in a massively multiplayer online role playing game (MMORPG). The outcome for the video games industry is to give the possibility to the developers to train mobs (non-playing characters) instead of programming them. This, of course, should drastically reduce development time and hopefully leads to more intelligent and more human like mobs. A MMORPG contains several different situations, we tackled the problem of choosing what to do (which spell/skill to use) and on which target in a setup where both allies and foes are present. We developed the model in Bayesian programming. This work resulted in a publication at MaxEnt 2010 .

The second work was on artificial intelligence for real time strategy (RTS) games. RTS games consist in controlling workers and armies to develop economy, technology, armies, and beat your opponent. Il involves strategy, tactics, and human skill (unit management). In 2010, we particularly worked on unit management with a Bayesian robotic programming model for unit control. Each unit is considered as a modal (scout, flock, fight, etc.) robot with sensory informations such as the topographical information (cliffs, buildings), its allies and foes positions and speeds (see figure ). The units group works as a hierarchical multi-agents system without inter-unit communication. We applied this work to Starcraft: Broodwar in the context of the AIIDE 2010 Starcraft AI Competition and it gave promising results.

Biochemical Probabilistic Inference is a new area of research which started in 2008 in close collaboration with LPPA-College de France and ProBAYES. In september 2010 Pierre Bessière moved to LPPA to boost this part of the work but stayed an associated researcher of E-motion.

Living organisms need to quickly react without waiting for a perfect evaluation of the consequences of their action. For instance, we perceive objects from retinal stimulation without the need for a complete knowledge of the underlying light-matter interactions. To account for this ability to reason with incomplete knowledge, it has been recently proposed that the brain works as a probabilistic machine, evaluating probability distribution over cognitively relevant variables. A number of Bayesian models have been shown to efficiently account for perceptive and behavioural tasks (see for a survey). However, little is known about the way subjective probabilities are represented and processed in the brain.

Numerous biochemical cellular signalling pathways have now been unravelled. These mechanisms involve the strong coupling of macromolecular assemblies, membrane voltage and diffusible messengers, including intracellular Ca2+ and other chemical substrates like cyclic nucleotides. Since transition between allosteric states and messenger diffusion are mainly powered by thermal agitation, descriptive models at the molecular level are also based on probabilistic relationships between biophysical and biochemical state variables .

Our proposal is based on the existence of a deep structural similarity between the probabilistic computation required at the macroscopic level to account for cognitive, perceptive and sensory motor abilities and the biochemical interactions of macromolecular assemblies and messengers involved in cellular signalling mechanisms. Our working hypothesis is then that biochemical processes constitute the nanoscale components of cognitive Bayesian inferences.

We propose 3 main novel ideas:

To be more specific, the starting ideas are the following:

The addressed scientific challenges imply a paradigm shift to a post Boolean computational approach that should have long term impact on different fundamental questions in both cognitive and ICT sciences: