**Computational geometry, Bayesian programming, Automatic learning, Automatized motions and actions, Space-time reasoning, Biologic inspiration, Robotics**

**Challenge:**
The project-team *e-Motion* aims at developing models and algorithms allowing us to build *``artificial systems''* including advanced sensori-motors loops, and exhibiting sufficiently efficient and robust behaviors for beeing able to operate in *open and dynamic environments* (i.e. in partially known environments, where time and dynamics play a major role), and leading to *varied interactions with humans*. Recent technological progresses on embedded computational power, on sensor technologies, and on miniaturised mechatronic systems, make the required technological breakthroughs potentially possible (including from the scalability point of view).

**Approach and research themes:**
In order to try to reach this objective, we propose to combine the
respective advantages of the *computational geometry* and of the *theory of probabilities*, while working in cooperation with neurophysiologists for trying to apply and experiment some *biological models*. This approach leads us to study, under these different points of view, three strongly correlated fundamental research
themes:

*Multimodal and incremental modelling of space and motion*. The
basic idea consists in continuously building (using preliminary knowledge
and current perceptive data) several types of models having complementary
functional specialisations (as suggested by neurophysiologists).
This leads us to address the following questions : how to model the various aspects of the real world ? how to consistently combine a priori knowledge and flows of perceptive data ? how to predict the motions and behaviors of the sensed object ?

*Motion planning for the physical world*. The main problem is to
simultaneously take into account various constraints of the physical world
such as non-collision, environment dynamicity, or reaction time, while
mastering the related algorithmic complexity.
Our approach for solving this problem consists in addressing two main questions :
how to construct incrementally efficient space-time representations ? how to define an iterative motion planning paradigm taking into account kinematics, dynamics, and time constraints ?

*Probabilistic inference for decision*. The problem to solve is
to be able to correctly reason about both the current knowledge of the
system and the associated uncertainties. Our approach for addressing this problem is to make use of the bayesian programming paradigm, by developping the related models and computational tools.

**Application domains: **
The main applications of this research are those aiming at introducing advanced and secured
robotized systems into the human ``living space''. In this context, we are focussing onto applications such as future cars and transportation systems, or service and intervention robotics (e.g. domestic tasks, civilian or military safety, entertainment). In cooperation with our spin-off company *Probayes*, we are also exploiting some spin-offs of our bayesian programming technique in application domains such as the diagnosis for the preventive maintenance of complex industrial plants, or the assistance to some financial decisions.

In spite of the significant technological advances made during the last decade, Robotics is still blocked by the problems of *scalability* and of the *real integration of robotised systems in our everyday life*. The reason for this stems primarily from the fact that models and technologies developed in the past (e.g. approaches based on logics, geometrical based methods, randomized search techniques, reactive architectures ...) have mainly reached their limits, and cannot be directly used for crossing the complexity gap introduced by the physical environments in which we are living. Indeed, such environments include complex multimodal data, are continuously changing and partially unpredictable, and generate complex interactions with humans. This means that *appropriate decision-making processes* taking into account these characteristics have to be designed, implemented and experimented in real situations. Such processes have to be efficient and robust enough for making it possible to meet the required reactivity characteristics, while beeing able to make appropriate decisions from complex and incomplete data and knowledges, i.e. by reasoning about a combination of partial a priori knowledges, of some incremental experimental data (including sensory data), and of some hidden variables. This means that new models and algorithms have to be designed for beeing able to formalise the intrinsic *``incompleteness''* of the problem, and to better model the intricate *``complexity''* of the real world.

The objective of the *e-Motion* project-team is to find formal and practical solutions to the previous unsolved problems, still very little addressed by our scientific community. Our bet is that our new approach based on both geometry and bayesian programming, will allow us to achieve the following technological breakthroughs:

*Motion autonomy in a dynamic complex world.*
We are especially interested in the problems arising from
the richness of the environments considered (i.e. how to model them
efficiently), from their dynamicity (i.e. taking explicitly into account
the ``space-time'' dimension), and from the large variety of possible
interactions (e.g. estimation and prediction of the behaviors of the
potential obstacles).

*Increased robustness of the automatic navigation processes.*
We put the emphasis on the problems of *incompleteness* (factors not taken into
account or hidden variables) inherent to the representation of any physical phenomenon. This dimension of the problem is generally empirically and
approximatively taken into account in traditional approaches, leading the
related systems to be poorly reliable. Our approach for dealing with this
problem is to convert the ``incompleteness'' into numerically quantifiable
data, coded in terms of probability distributions and referred to as *``uncertainties''*. Then, such random variables can be combined, evaluated,
and used in various decision-making processes.

*Intuitive programming of artificial systems and of their associated reactive processes.*
Our approach consists in using as far as possible *learning* processes
(supervised or not), in order to be able to combine the a priori knowledge
(called ``preliminary data'') and the past experience of the system (called
``experimental data''); this approach should permit us to gradually obtain
systems more robust and better adapted to the problems at hand. We will use and generalize our new concept of *Bayesian Programming* for developing the required processes.

As previously mentionned, the main applications of our research are those aiming at introducing advanced and secured robotized systems into the human ``living space''. In this context, we are focussing onto the following application domains :

*Future cars and transportation systems.* This application domain
should quickly change under the effects of both new technologies and
current economical and security requirements of our modern society. Various
technologies are currently studied and developed by research laboratories
and industry. Among these technologies, we are interested in *ADAS**Automatic Driving* functions allowing fully automatic displacements of private or public vehicles in particular driving situations and/or in some equipped areas (e.g. automated car parks or captive fleets in downtown centres).

*Service and intervention robotics.* This application domain
should really explode as soon as robust industrial products, easily usable
by non-specialists, and of a reasonable cost will appear on the market. One
can quote in this field of application, home robots (such as for example
current vacuum cleaning robots which are both too expensive and poorly
efficient), active surveillance systems (e.g. surveillance mobile robots,
civilian or military safety, etc.), entertainment robots (such as for
example the Sony robots *Aibo* or *Qrio*), or robotised systems for
assisting elderly and/or to disabled people. The technologies we are
developing should obviously be of a major interest for such types of
applications.

*Potential spin-offs in some other application domains.* The
software technologies we are developing (for bayesian programming) should also
have a potential impact on a large spectrum of application domains, covering fields
as varied as the interaction with autonomous agents in a virtual world (e.g. in the
video games), the modelling of some biological sensory-motor systems for
helping neurophysiologists to understand living systems, or applications in
economic sectors far away from robotics like those of finance or plant
maintenance (applications currently covered by our start-up *Probayes*
commercializing products based on Bayesian programming).

*ColDetect. People involved : Christian Laugier, Kenneth Sundaraj*.
This library has been implemented for providing robust and efficient *collision detection, exact distance computation, and contact localisation of three-dimensional polygonal objects*. These objects can be concave or convex, rigid or deformable. The library is numerically robust, i.e. the algorithm is not subject to conditioning problems and requires no special handling of nongeneric cases. ColDetect has been implemented in standard C++ and relies heavily on STL in order to be as fast and memory efficient. Currently it compiles under GNU g++ version 2.95 and 3.2. *ColDetect* is patented under the french APP patent
#IDDN.FR.001.280011.000.S.P.2004.000.10000.

*ProBT. People involved : Juan-Manuel Ahuactzin, Kamel Mekhnacha, Pierre Bessière, Emmanuel Mazer, Manuel Yguel, Olivier Aycard.*
Formerly known as *OPL*, *ProBT* is a C++ library for developing efficient
Bayesian software. It is available for Linux, Unix, PC Windows (Visual
C++), MacOS9, MacOSX and Irix systems. The ProBT library has two main
components: (i) a friendly Application Program Interface (API) for building
Bayesian models, and (ii) a high-performance Bayesian Inference Engine (BIE)
allowing to execute all the probability calculus in exact or approximate
way. *ProBT* is now commercialized by our start-up *Probayes*; it represents the
main Bayesian programming tool of the *e-Motion* project-team, and it is
currently used in a variety of external projects both in the academic and industrial field.

*Cycab Simulator. People involved : Cédric Pradalier.* In order to perform pre-test and to provide help for Cycab developers, a
Cycab simulator has been developed. This simulator is intended to
simulate hardware and low-level drivers, in order to produce a temporal
behaviour (refresh frequency, scheduling...) similar to what can be found
on the Cycab. Furthermore, a hierarchy of C++ classes has been developed
in order to keep a consistent interface between the simulated CyCab and the
real one. Applications written and tested on the simulated robot can then
be settled to the real one with only minor modifications (instantiating one
class or the other). Sensors and environment are also simulated, so that
complete applications can be developed on this test bed. Finally, we also
provide developer with an TCP/IP controllable Cycab, consistent with
simulated and real Cycab in term of C++ interface.
The Cycab Simulator is currently widely used by the researchers of the *e-Motion* project-team; it has also been used in a collaboration with the RIA team (LAAS, Toulouse), and with the LAG laboratory (IMAG, Grenoble). A recent extension of the system is also used for student home-work in the scope of robotics courses (Summer school on image & robotics 2004, and Master IVR at ENSIMAG Grenoble 2004-05).

*VisteoPhysic. People involved : Cesar Mendoza, Kenneth Sundaraj, Christian Laugier.* This library provides efficient tools for deformable object simulation. It includes the Finite Element Method (FEM) and the Long Element Method (LEM) deformable models for physical simulation. It also has interactions models for collision detection, exact distance computation, and contact localization of three-dimensional polygonal objects. These objects can be concave or convex, rigid or deformable. This library is numerically robust - the algorithms are not subject to conditioning problems, and requires no special handling of nongeneric cases. *VisteoPhysic* has been implemented in standard C++ and relies heavily on STL in order to be as fast and memory efficient. The library was developed in collaboration with XL-Studio and is patented under the french APP patent #IDDN.FR.001.210025.000.S.P.2004.000.10000

*VDM. People involved : Kenneth Sundaraj, Christian Laugier*. This library provides efficient tools for real-time simulation of biological tissue. It is based on a new physical model called the Volume Distribution Method (VDM). This physical model is based on Pascal's principle and uses volume conservation as boundary conditions. This software is distributed freely on the internet under a GPL license. Currently, this software has been used in a prototype system of an echographic thigh exam simulator and an arthroscopy knee reconstruction simulator for the ACL ligament. This work was within the framework of a collaboration with Aesculap-Bbraun (Navigation systems for surgery). VDM version 1.0 has a french APP patent #IDDN.FR.001.280012.000.S.P.2004.000.10000

Simultaneous Localization And Mapping (SLAM) is a well known problem in the robotic community as long as static environments are concerned. The problem of SLAM involves incrementally building a map of its environment as an autonomous robot localizes itself with respect to the map it had built. Since 2002, we have demonstrated our ability to perform localization and mapping in such an environment.

To extend our knowledge on this field, we initiated this year a research thread on localization and mapping in a changing environment. We consider the case of a CyCab robot equipped with a Sick laser range finder evolving on a car park with occupied and free parking lots. With this sensor in such an environment, the robot can only sense cars, either parked or moving. We have a three step objective: firstly to extract hypothesized cars from sensor output; secondly, to use these hypothesized cars to build a car map of the parking area, i.e. to build a map of parked car, being able to detect when a new car is parked or when a parked car left; thirdly, using static cars as landmarks, we can refine our position estimation in the car park and estimate other cars' movement.

The first part of this work was conducted last year by a master student (J. Lorieux, ). What was obtained was the static extraction of vehicle hypothesis from laser range data which constitutes the first stage of the vehicle hypothesis filtering process.

This year, the second objective was achieved by a second master student (C. Tay Meng Keat, ). We managed to improve the robustness of the vehicle extraction by performing strict double checking during the second stage of vehicle hypothesis filtering.

Figure (left) shows hypothesized cars (rectangles) extracted from a laser scan, whereas Figure (right) shows remaining hypotheses after our improved filtering method.

In order to build a map of the car park, we wanted to implement SLAM techniques, using the extracted hypothesised vehicle as landmarks. Amongst existing SLAM techniques, we chose to use the method FastSLAM . The idea behind this method is that knowing the full trajectory of the robot, the state of the landmarks are conditionnally independant. This leads to computational simplication and consequently a fast and efficient algorithm, for which it is particularly suitable for realtime applications.

In FastSLAM implementations, the probability distribution on the robot state is represented by a cloud of particles (as in particle filters ). The specific characteristic of FastSLAM is that each particle contains its own estimation of the map, i.e. the estimation of the map knowing the full trajectory since the start. Consequently, when we need a usable representation of the map (for path planning for instance), we have to fuse all the map hypotheses contained in all the particles. This year, we have proposed an algorithm to implement, in a quite robust way, this fusion process allowing the construction of the final map .

These results have been submitted to the ICRA'05 conference. Furthermore, C. Tay Meng Keat is beginning a PhD thesis centered on the simultaneous localization and mapping of a dynamic environment, which corresponds to the third objective.

This work concerns the hierarchical and bio-inspired modeling of navigation capacities, using the Bayesian formalism. Our main previous contribution in this domain is a framework developped during Julien Diard's PhD, called the Bayesian Map formalism .

In 2004, we have been interested in two domains of application of this model:

*Robotics*. We have first argued that the Bayesian Map formalism is a possible mariage between, on the one hand, ``bio-inspired models'' (which are a promising tool for developping flexible navigation skills in mobile robots) and, on the other hand, ``hierarchical solutions'', which are needed to model larger environments than what is possible in state-of-the-art control architectures. The Bayesian Map formalism we have developped has been also successfully applied on a Koala robot during proof-of-concept experiments. For instance, we have demonstrated that it was possible to use Bayesian inference as a way to model high-level localization as a ``recognition'' of lower-level sensory-motor models . We are also currently developping a full-scale control architecture using the Bayesian Map formalism, including the learning of the sub-models as well as of the way sub-models are articulated . We also aim at integrating the notion of hierarchical and bio-inspired models of navigation in the context of probabilistic path planning .

*Cognitive modeling*. Part of this research track has been done in collaboration with the Laboratory of Physiology of Perception and Action, Collège de France, Paris, during Julien Diard's post-doctoral visit there. In this domain, we are arguing that the Bayesian Map framework is a promising tool for modeling human navigation skills . Indeed, on the one hand, it is based on the probability calculus which helps treat uncertainties and decision making in a mathematically sound manner. On the other hand, it offers modeling tools for modular and hierarchical systems, which probably are the best tools for understanding human navigation as a sum of sub-systems (for instance, sub-systems that deal with parts of the environment, part of the sensory input, part of the navigation task to solve, etc). The Bayesian Map formalism proposes mathematical operators for modeling the way such sub-systems can be articulated. For instance, we have developped a Bayesian model of the coordination between two cognitive strategies (landmark-based navigation and path integration) in humans involved in a navigation task in a virtual environment . This model qualitatively describes errors made by human participants when cues used for memorising paths are removed and missing for the reproduction of the path. Current research is aimed at precising the model, in particular at modeling the influence of geometrical and visual landmarks in path memorisation.

To navigate or plan motions for a robotic system placed in an environment with moving objects, reasoning about the future behaviour of the moving objects is required. In most cases, this future behaviour is unknown and one has to resort to predictions. Most prediction techniques found in the literature are limited to short-term prediction only (a few seconds at best) which is not satisfactory especially from a motion planning point of view.

In 2003, we have started to explore the problem of medium-term motion prediction for moving objects. As a result, we have proposed a novel cluster-based technique that learns typical motion patterns using pairwise clustering and use those patterns to predict future motion .

During 2004, we have developed a number of libraries and utilities (environment modelling, sensor modelling, camera calibration, communications) that will be used for further experiments. On the theoretical side, we have started to work on a new technique whose purpose is to address a number of issues that were not solved by our first approach:

*Prediction of unobserved patterns:* our first approach was
not able to address a situation where a moving object would ``change
its mind'' and switch from one typical motion pattern to another one.
The new approach which relies upon the notion of goals rather than
typical motion patterns will be able to address this kind of problem.

*On-line/adaptive learning:* the clustering technique used in
our first approach is an off-line process taking as input a data set of
observed trajectories. The new learning technique will be able to learn
on-line, *ie* to take into account newly observed trajectories
while carrying out the prediction.

The approach we are developing operates at two levels:

*State Representation:* space sampling should be dense in
interesting areas and coarse in areas where nothing happens. For this, we
are working on discretisation using a self-organised map (SOM) known as the
Growing Neural Gas . One of the main advantages of this
approach is that it is adaptive and works on-line
(Fig. ).

*Object Motion Modelling:* motion is modelled as being
dependent on a plan, the plan itself is hierarchically decomposed into
subgoals which, at the lowest level, correspond to probabilities of
transition between states. The specific model we are focusing on is the
Abstract Hidden Markov Model . Thanks to the chosen state
representation, it is possible to learn transition probabilities at the
same time that we are constructing the state representation
(Fig. ).

Unlike regular cruise control systems, Adaptive Cruise Control (ACC)
systems use a range sensor to regulate the speed of the car while ensuring
collision avoidance with the vehicle in front. ACC systems were introduced
on the automotive market in 1999. Since then, surveys and experimental
assessments have demonstrated the interest for this kind of systems. They
are the first step towards the design of future Advanced Driver Assistance
Systems (ADAS) that should help the driver in increasingly complex driving
tasks. The use of today's commercially available ACC systems is pretty
much limited to motorways or urban expressways without crossings. The
traffic situations encountered are rather simple and attention can be
focused on a few, well defined detected objects (cars and trucks). A wider
use of such systems requires to extend their range of operation to more
complex situations in dense traffic environments, around or inside urban
areas. In such areas, traffic is characterized by lower speeds, tight
curves, traffic signs, crossings and ``fragile'' traffic participants such
as motorbikes, bicycles or pedestrians. A prerequisite to a reliable ADAS
in such complex traffic situations is an estimation of the dynamic
characteristics of the traffic participants, such as position and velocity.
This problem can be seen as a *Multi-Target Tracking* problem.
Classical approaches of Multi-Target Tracking are designed for military
applications and therefore do not answer
urban driving specificities. Numerous methods (JPDA, PMHT, etc.) consider
known and constant number of targets. Other methods (MHT) allow
the creation of new tracks, but they are intractable in situations
involving numerous appearances, disappearances and occlusions of a large
number of rapidly manoeuvering targets.We have chosen to express the problem
of environment representation in a different way. We prefer to estimate the
occupied and free space of the environment of our vehicle in a 4 dimensions grid,
including 2D-positions and 2D-velocities. We called the
corresponding model the *Bayesian Occupancy Filter (BOF)*. This model is inspired by *occupancy grids* which
have been extensively used for mapping and localisation in static
environments.

In 2003, to demonstrate the tractability and the relevance of the BOF, a collision avoidance behavior in a dynamic environment has been implemented on the Cycab robot .

These experimentations have shown that the main problem of mapping through
*bayesian occupancy filter (BOF)* is the amount of data structure it needs.
In fact this approach deals with dense information in 4 dimension. So all
algorithms using this kind of data representation have to work on small areas.
To avoid this restriction we work on a new representation of this data using
multiscaled time-space representations like wavelets (Fig. ).
We assume that urban area are localy very regular at men scale. By example,
walls surround buildings which have often regular shapes and zero velocity, and
roads separate them. So, we represent them with patches of free space for roads and
occupied space with zero velocity for buildings. In this terms we can filter 4D map
data to resume occupancy space with 4D standard patches at different scales. We only
work with a list of patches instead of raw cells from discretized high dimensionnal
space.
This technique is under test in real urban environments with the puvame project.
The Cycab robot is used to simulate bus motion in crossroads with pedestrian
displacements and parked vehicles which could hide each other.

One issue in the modelling of space is the perception of its structure. *Structure from Motion* (SfM) is a way to extract some geometrical information
from the optic flow generated by the movement either of an object or the
observer ( ). Sfm is a common field of interest to both
computer vision and experimental physiology.
Within the framework of the BIBA project, we have investigated this year the
psychophysical findings on human perception of shapes from motion. We were
particularly interested in the optic flow of a plane that presents several
amiguities. Recovering geometry from optic flow is then an inverse and
ill-posed problem. We proposed in a bayesian program to take
into account different results integrate them into a single model. So far we
were able to reproduce the results of five psychophysical experiment such as
the difference between active and passive performance or the shear or (winding
angle) effect described in .
Thus bayesian modelling has been an adequate framework to deal with this
inverse and ill-posed problem

This year, in this research topic related to the previous Sharp project, we have contributed in the following:

Soft Tissue Simulation

Bayesian Collision Detection

Computer Aided Surgery

We have completed the development of our new physical model for soft tissue simulation. This model, called the Volume Distribution Method (VDM), has been implemented and packaged into a freeware library under an APP patent. The VDM model is derived from bulk variables like pressure and volume. In terms of complexity, this new model is one order of magnitude lower than classical volumic models like the Finite Element Method (FEM) . Pascal's principle and volume conservation are used as boundary conditions. Since volume conservation is observed in soft tissue, the VDM model is a suitable choice. Furthermore, its lower complexity makes it an interesting alternative for interactive-time applications.

On the past, the collision detection problem has been tackled from the
exact point of view. That is, given the configurations of two
geometric entities, the response of a collision detection algorithm
is: ``in collision'' or ``not in collision''. In the *e-Motion* team, we
are exploring the collision detection problem from the probabilistic
point of view. In this case, the response of the collision detection
algorithm is: ``the probabillity of collision is q'' where q belongs
to the interval [0, 1]. The main advantage of this approach is
speediness. In effect computing q only requires of a few of
multiplications.

Numerous Virtual Reality or Robotics applications do not require of an exact collision detection. We believe that this schema can be used for developing applications where (i) the collision detection does not have to be exact (such as 3D video games) and (ii) the fast yes/not collision detection answer is required (e.g. in dynamic situations).

In order to compute the ``probability of collision'' between two entities, we are using
a sensor fusion scheme. The fusion of n sensor distributions
computes the probability of the collision variable. The sensor
distributions are computed in a off-line learning phase, and stored in
the mesh data base. When an application loads a set of meshes, the
sensor distributions are also loaded. The Bayesian detection algorithm
uses this data to compute the probability of collision for a given set
of pairs.

Our preliminary experimental results shows that this approachs is
promising. Here we depict one of our experiments inspired from motion
planning. It consists in simulating the fall of the stick on
a torus. At the beginning the stick is placed at the
center top position of the bounding sphere of the torus. At each time
iteration the position on z is decremented of z while x
and y remains constant. The orientation is chosen (with our
algorithm) so that the probability of collision is minimized. These
minimization is executed in a vicinity of the configuration at time
(t-1). The result is shown on Figure . As it can be
seen, the stick is oriented so that it avoids the collision while
crossing the center of the torus.

Our preliminary results are very encouraging. In effect, if we
consider that two configurations are in collision if pp_{0} then
the obtained presicion is p_{0}. The running time of the Bayesian
program is the time required to realizing a few multiplication, as
much as the double of the number of virtual sensors.

In this context, we have worked in computer aided surgery for the Total Hip Prosthesis (THP) in collaboration with Aesculap-BBraun. The aim of the project is to obtain a CT/MRI-free 3D volume generation of the hip for preoperative planning and navigation, and operative guidance for a normal hip. The approach is to obtain an approximation of the real 3D hip volume of the patient by deforming a nominal 3D hip mesh (Atlas). The deformation consists in driving a set of control points on the hip mesh that will match specific positions provided by 2D radiographic data and 3D echographic data of the patient. Our goal is to use the obtained 3D volume for the preoperative planning and navigation. In the preoperative planning the 3D hip volume will allow the surgeon to select the prosthesis type(size and form) and position. The position will be for a normal hip rang of motion and to avoid prosthesis collision. The 3D volume will be used to guide the surgeon in the prosthesis implantation providing the 3D position and orientation of the patient's hip and the surgical tools.

First evaluations have been made obtaining the 3D coordinates of a
``saw bone model'' with the ``orthopilot system'' of Aesculap-BBraun. The deformation have
been made using the 2D coordinates of the corresponding radiographic
data and the 3D echographic data. The error was calculated between
the real 3D coordinates of the sawbone and the deformed hip mesh. We
have obtained a mean error of about 1mm for the 3D obtained hip
volume. Currently work aims at improving accuracy of the 3D obtained
hip volume (more specially in the acetabular region), and at completing the experimental validations (Fig. ).

Two methods developed last year were integrated together this year: On the first hand, we used Bayesian programming to design and develop an efficient and reliable reactive obstacle avoidance system. On the second hand, we developed and experimented the integration of localization and mapping, path planning and trajectory execution. This year, in collaboration with the SED service of INRIA Rhône-Alpes, we integrated obstacle avoidance in our trajectory execution module. This integration stressed the fact that the control law using the CyCab flatness property is very sensitive to perturbations. So, using again Bayesian programming, we had to design a trajectory tracking behaviour that can be used alternatively when flat control law is not reliable any longer. This work has been published in the ICRA'04 conference , and a more complete version of the paper has been accepted for publication in the Robotic And Autonomous System journal .

Autonomous behavior is also possible whithout mapping, and in this direction a programming framework based on the Bayesian paradigm was proposed in for conception and development of an autonomous behavior for a mobile robot evolving in an unmodified environment. This framework assumes a desired behavior is specified and in order to deal with resulting complexity under limited resources, a general structure based on cascading Bayes Filters is described, applying some insights of natural systems. Some specific questions are addressed, as the role of attention in perception and how action selection influences perception tuning. Practical implementation of this for an animal-like behavior is under development and the reactive obstacle avoidance method presented above is employed in order to achieve safe execution of infered motor commands.

The purpose of this research work is to study how the Markov Decision
Process Theory can be used to robustify mobile robot navigation.
To reach a given goal, a mobile robot computes a motion plan first,
(*ie* a sequence of actions that will take it to its goal), and then
executes it. Markov Decision Processes (MDPs) have been successfully used
to solve these two problems. Such approaches use a graph representation of
the robot's state space and their main advantage is that they provide a
theoretical framework to deal with the uncertainties related to the robot's
motor and perceptive actions during both planning and execution
stages. Unfortunately, their algorithm complexity is exponential in the
number of edges of the graph which limits their application to complex
problems.

Our approach is to address this complexity issue by reducing the number of states through aggregation techniques. We have proposed a MDP-based planning method that uses a hierarchical representation of the robot's state space (based on a quadtree decomposition of the environment). Besides, the actions used better integrate the kinematic constraints of a wheeled mobile robot. These two features yield a motion planner more efficient and better suited to plan robust motion strategies .

In 2004, we have focussed on the execution stage. An execution scheme based upon Markov Localisation has been defined and validated experimentally. Central in this validation is the transition function (that encodes the uncertainties related to the robot actions), and the sensors' model: a learning procedure using the real robot and its sensors have been defined and experiments with a Koala mobile platform has demonstrated the robustness of the whole navigation approach (Fig. ). A paper presenting these results has been submitted to the IEEE Int. Conf. on Robotics and Automation 2005.

To navigate in open environments, that is to say where there is a lot of moving obstacles (pedestrians, bicycles, cars, ...), we need to detect obstacles (moving or not) and to estimate their state (position, speed, ...). Actually, today's techniques do not allow to measure obstacles motion.

That is why in this work we proposed a new method to compute obstacles position and speed. To be applied to a real robot in dynamic environments we have a real-time constraint.

In a first part, we proposed a new method to compute optical flow in real-time that allows to evaluate a confidence measure. It has been experimentally shown that this algorithm works at the frame rate of 15-30 Hz, and is auto-adaptative to various scene conditions.

In a second part, we have proposed an obstacle detection technique based on the optical flow. We use it to compute the real motion of the obstacles in the image, and to extract them by comparing their optical flows to the optical flow on the ground. By extracting the obstacles, and knowing their optical flows, we can compute their position and speed.

Experiments have been performed on the Cycab platform on top of which is fixed a stereo camera (only one of the two cameras have been used).

An paper about this work has been submitted to ICRA 2005. Current work (Christophe Braillon's PhD thesis) focus on navigation in open and dynamic environments, using optical flow data combined with some other sensing data and a priori knowledge.

When placed in a dynamic environment, *ie* an environment with
moving objects, a robotic system cannot stand still since it might be
collided by one of the moving objects. In such a framework, a real-time
constraint is imposed upon the navigation or motion planning system of the
robot considered: it has a limited time only to determine its future course
of action. The time available is a function of what is called the
*dynamicity* of the environment which is directly related to the
speed of both the moving objects and the robotic system.
Previous works dealing with robotic systems placed in dynamic environments
fall into three families:

*Motion planning approaches* wherein a complete motion to the
goal is computed . Unfortunately, the intrinsic complexity of motion
planning restricts the use of this kind of approach to simple robotic
systems only and environments with low dynamicity.

*Reactive approaches* wherein the action for the next time-step
only is computed . Such approaches
do verify the real-time constraint. Unfortunately, their lack of
lookahead raises the convergence and safety issues. In other words, what
guarantee do the robot have to ever reach the goal and to never find
itself in a critical situation?

As a partial answer to the convergence and safety issues,
*extended reactive approaches* have been proposed wherein the
reactive scheme is extended over a few time-steps .

Our own answer to the problem at hand is the *partial motion
planning approach*. It is similar to an extended reactive approach in the
sense that only a partial motion is computed. However, instead of
extending a reactive approach over a few time-steps only, it is proposed
instead to use randomised motion planning techniques so as to benefit from
the advantages of such techniques and to carry out the planning effort as
far as possible in the future given the time available.
Of course, since only a partial motion is computed, it is necessary to
iterate the partial motion planning process until the goal is
reached. Accordingly, the whole approach is named *Iterative Partial
Motion Planning* (IPMP). Fig. depicts the IPMP cycle.

As a matter of fact, the iterative nature of IPMP is doubly required when
the robotic system at hand is placed in a uncertain dynamic environment,
*ie* an environment for which everything is not known in advance (in
particular, the future behaviour of the moving objects).
Motion planning means reasoning about the future. When the future is
unknown, one has to resort to predictions, predictions whose validity
duration is limited in most cases. An iterative planning scheme permits to
take into account the unexpected changes of the environment by updating the
predictions at a given frequency (which is also determined by the
environment dynamicity)..

Like the reactive and extended reactive approaches, IPMP has to address the
convergence and safety issues.
As far as the second question is concerned, the unrealistic convergence
conditions established in leave little
hope (it is hardly surprising if the system is placed in an environment
with no a priori information about the moving obstacles).
The safety issue is more important and we would like some guarantee that
the system never finds itself in a critical situation. To address this
safety issue, we have been exploring two different solutions: the first one
relying upon the concept of *Inevitable Collision States* and the
second one upon the concept of *Non Linear Velocity Obstacles*
(*cf* sections and respectively).

In 2004, a complete prototype of IPMP using the Inevitable Collision States has been implemented and tested for a car-like vehicle (Fig. ). This work yielded a paper that has been submitted to the IEEE Int. Conf. on Robotics and Automation 2005. In 2005, it is planned to plug the IPMP module on a Cycab experimental vehicle. To do so, the coupling between IPMP and the control level of the Cycab will be studied (Ph.D topic of Stéphane Petti). Live experiments should validate the whole approach.

In 2002, we introduced the novel concept of *inevitable collision
states* (ICS). In general, an ICS for a given robotic system can be
defined as a state for which, no matter what the future trajectory followed
by the system is, a collision eventually occurs with an obstacle of the
environment.

An inevitable collision state takes into account the dynamics of both the robotic system and the obstacles, fixed or moving. This very general concept can therefore be useful both for navigation and motion planning purposes (for its own safety, a robotic system should never find itself in an inevitable collision state).

To illustrate the interest of this concept, it was applied in 2002 to a
problem of *safe motion planning* for a robotic system subject to
sensing constraints in a partially known environment (*ie* that may
contain unexpected obstacles) .
In safe motion planning, the issue is to compute motions for which it is
guaranteed that, no matter what happens at execution time, the robotic
system never finds itself in a situation where there is no way for it to
avoid collision with an unexpected obstacle.

While 2003 was dedicated to further the exploration of the ICS concept by demonstrating a number of properties are fundamental for their characterisation . 2004 was mostly spent studying the use of the ICS concept to guarantee safety within the IPMP scheme. We have established that a partial trajectory is not within ICS iff it is collision-free and its final state is not an ICS. This simplifying property has permitted to improve the performance of IPMP, it is presented in a paper that has been submitted to the IEEE Int. Conf. on Robotics and Automation 2005.

Work on the NLVO paradigm (Non-Linear Velocity-Obstacles) has been initiated in 2001 , while looking for a way to improve the decision process of a robot (service robot, intelligent transportation system, virtual character...) moving at high speeds in a partially unknown environment amidst static and mobile objects. The NLVO concept is inspired from the V-obstacle concept. It aims to compute at the same time, the time to collision for all admissible movements of the robot. This information allows a fast estimation of the risk of collision associated with any feasible movement of a robot . This information is a high level representation of the environment in the velocity space of the robot. Contrary to non classical representations, one can stress that such an approach does not loose any possible solutions and takes into account the dynamics of the world.

Two navigation methods have been derived from the NLVO: the first method is reactive, it computes only the next move of the robot: It is more suitable for fast changing environments like a crowd, but may cause the robot to be blocked at some points. The second method aims at avoiding this drawback by building a complete trajectory to the goal. The solution is searched in a tree, iteratively built so that the planning can be suspended at any time to return an intermediate result. This ensures a certain reactivity of the robot, needed for its safety. NLVO are used to expand the tree while the branches to explore are chosen by a heuristic function aiming at optimising the travelling time. The computed trajectories are safe (due to NLVO). They also limit the cases that prevent the robot from reaching its goal, remain coherent between two successive decisions of the robot and stay away from unpredicted objects (thanks to the algorithm used to update the tree).

Computing a NLVO requires a priori knowledge about the future behaviour of
the moving objects. In 2004, we have worked on the coupling between the
NLVO reactive approach and our moving objects' future motion prediction
function (*cf*
section ).

Part of our work this year consisted in developing a perceptual servoing on a sensori-motor trajectory. We call ``sensori-motor trajectory'' a trajectory, which instead of being described as sequence of robot configuration or position, is described as a sequence of exteroceptive perceptions (sensors) associated with proprioceptive perceptions (motors). Our perceptual servoing consists in being able to follow such a trajectory, starting from any configuration in its neighbourhood, going backward or forward, and being able to divert from the trajectory for obstacle avoidance. This task has been addressed in the Bayesian framework, and decomposed in four subtasks: error estimation (i.e. relative localization), trajectory tracking, obstacle avoidance and online estimation of the observation consistency with respect to the localization hypotheses. Successful experimentations were performed, in simulation and on the real robot on the car park of INRIA Rhône-Alpes, showing the validity of the approach. This work has been published in ICRA'04 and ISER'04. A more complete article, based on the results presented in C. Pradalier's PhD thesis has been submitted to IEEE transation on robotics.

Information sensed by a robot evolving in a real environment is both too large and uncertain for being easily exploited by the navigation function and the related reactive behaviors.

The first problem is to extract relevant information for the task that the robot is currently performing. Usually this extraction is done by a human programmer who, for instance, decides that proximity sensors are relevant for an obstacle avoidance behavior. But a really autonomous robot should be able to discover that kind of information, which is called ``preliminary knowledge'' in our bayesian programming framework. Some results have been obtained using methods based on feature selection to determine relevant sensors for a tracking task .

This year we applied these technics to auto-supervised learning, a compromise between supervised and completely unsupervised learning, consisting in relying on previous knowledge to acquire new skills .

On top of that, we developed a genetic algorithm framework for learning another kind of prior knowledge (``parametrical forms''). This was done considering that animals got their prior knowledge by natural selection. The main idea is to define an entropy driven evolution. This work leads to the implementation of several visual behaviors on the BIBA robot.

The second problem consists in combining several behaviors inspired by biology, in order to perform robustly a variety of tasks. Some interesting results using Bayesian programming have been obtained and experimented on the BIBA robot (for obstacle avoidance and homing strategies) . In relation to reactive behavior, a programming framework based on the Bayesian paradigm was proposed in for conception and development of an autonomous behavior for a mobile robot evolving in an unmodified environment. In order to deal with complexity of desired behavior under limited resources, a general structure based on cascading Bayes Filters is described, applying some insights of natural systems. Questions as attention in perception and action selection influences in perception tuning are addressed, and practical implementation of the framework proposed are in execution. Once motor commands are infered, the reactive obstacle avoidance method presented above is employed in order to achieve safe execution of commands.

Bayesian programming techniques used in robotics can be applied to synthetic characters such as those in video games (see figure ). Both game play and development methodologies can benefit from the learning techniques that arise naturally from these bayesian models.

We have pursued work on developing bayesian programs for behaviour selection and execution for video game characters. Previous results (See for a synthesis on using bayesian programs to execute and learn automata-like behaviours.) have shown that these techniques match or exceed the capabilities of common methods used to control characters. We are now focusing on learning behaviours by demonstration, which includes developing methods to identify human playing behaviours .

Two new components have been integrated in the inference algorithms of
ProBT Bayesian Inference Engine (BIE): (i) an *expectation
maximization* algorithm and (ii) a method for estimating probability
distributions of discrete variables by using the *entropy maximization
principle*.

Learning using incomplete data is a central issue in Bayesian reasoning. Indeed, some variables of the model may be unobserved and or/some value are missing for another set of variables.

The EM algorithm (Expectation-Maximization) is a general-purpose algorithm used to learn (estimate) model parameters using incomplete data. The idea behind the EM algorithm is intuitive and natural and it has been formulated and applied to a variety of problems. The basic idea of EM is to associate with the given incomplete-data problem, a complete-data problem for which Maximum Likelihood (ML) (or Maximum a posteriori (MAP)) estimation is computationally more tractable.

The EM algorithm consists in reformulating the incomplete-data problem in terms of this more easily solved complete-data problem. The E-step consists in producing data for the complete-data problem, using the observed data set of the incomplete-data problem and the current value of the parameters. The complete-data is then used to fit the free parameters (by maximizing the Likelihood or the data).

Given a probabilistic model, using the EM algorithm to learn the free
parameters of its distributions consists in estimating the joint
distribution over all variables of the model using the observed data on a
subset of them. Suppose, for example, that we are interested in estimating
the distribution of a given variable X as a mixture of Gaussians. Given the
number (n) of the kernels (Gaussians) of the mixture, the problem is to
estimate the distribution of the class variable C and the n mean and
variance values of the corresponding Gaussians. In other words, we have to
estimate the joint distribution P(C)P(X|C) where P(C) is a probability
table over C and, for each value c of C, P(X|c) is a Gaussian
distribution.

In the ProBT BIE, the E-step is simply considered as an inference problem
consisting in constructing the probability table P(C|x) over C for each
observed value x of X. In the M-step, the incomplete-data (x) is replaced by
n complete-data (c1, x), .., (cn, x) weighted by the corresponding
probability values P(C = c1) P(C = Cn). This complete-data set is then used to
estimate the gaussians G1(X) Gn(X). This estimation is made easy thanks to
the ProBT complete-data learning module (estimation of histograms, 1d and
nd-gaussians, )

The algorithm of *entropy maximization* calculus was integrated in the ProBT
BIE as an optimization problem. The ProBT API allows the observable
encoding. This scheme is particularly useful when the constraints
specification is preferred to probability distributions specification.

In order to respond to computationally intensive ProBT applications exceeding the computing power of a sequential machine, we are currently developing a parallel version of the ProBT BIE. The parallel version is designed in collaboration with the APACHE project. Two specialized environments for parallel probability calculus will be included in ProBT:

**Symmetric Multiprocessors (SMP)/ Thread computing.** In this
environment 2 to 64 processor in a single computer share all the
resources (memory, bus, I/O system) to run the ProBT BIE. Each processor
handles a specific part of the probability calculus. Once these calculus
are finished the partial results are recovered by the master processor
who returns the final result.

**Cluster computing.** In this environment a copy of the ProBT BIE
runs on each of the workstations (nodes) conforming the cluster. Each
node handles a specific part of the probability calculus. In addition,
each of the nodes can be a single or multiprocessor system (SMP). The set
of nodes work together as a single high-speed computer. The advantage
with respect to other parallel computer architectures is that a cluster
uses standard technology accessible at lower cost.

The ProBT API is the same for sequential or parallel computing. That is, a program specified in the ProBT API doesn't change. Just an additional initialization call function is required for the parallel environments (SMP or cluster).

Learning from data is a central issue when using Bayesian methods. This learning may concern the estimation of the free parameters of probability distributions of a Bayesian model, and/or the dependencies between the variables of the model.

For the first problem (i.e. distributions free parameters estimation), numerous classes allowing to learn standard probability distributions have been developed this year.

Future work in this project will concern the second problem (i.e. learning variables dependencies possibly under a set of constraints).

CyberCars [August 2001-July 2004]

European project IST-2000-28487 CyberCars, "Cybernetic Cars for a New Transportation System in the Cities", (http://www.cybercars.org).

The goals of this project are the development and experimentation of
new techniques of transport. These techniques are based on the use of
individual and automatical vehicles which circulate in the streets of
the cities or private sites instead of using of a private car complementary to public transport. The CyberCars consortium includes 14 partners coming from industry and public research. The contribution of *e-Motion* in CyberCars relates to driving automation.

Carsense [January 2000-December 2003]

European project IST 1999-12224 CarSense, ``Sensing of Car Environment at
Low Speed Driving'' (http://www.carsense.org).
A consortium of 12 european car manufacturers, suppliers and research institutes are together under the head of the CARSENSE programme. This programme, sponsored by the EC shall develop a sensor system, that shall give sufficient information on the car environment at low speeds in order to allow low speed driving. This project includes european industrials from car industry
(Renault, BMW, Lucas Varity, Thomson Detexys, Ibeo, etc.) and research
institutes (Inria, Inrets, Livic). *e-Motion* is in charge of the data fusion subject.

Profusion [2004-2008]

European project, PreVENT Programme (Preventive and Active Safety Applications) Profusion, ``Project for Robust and Optimized Perception by Sensor Data Fusion''

The first phase of ProFusion was a short six months project, which began
February 1^{st} and ended July 31^{st}. It was the first stage of a
horizontal activity within IP PReVENT, that aimed at: establishing a
forum including representatives of specified vertical subprojects for
exchanges on topics related to sensors and sensor data fusion,
circulating, feeding back and synthesizing information exchanged, and
exploiting the outcome from these exchanges to specify and propose one
or more new horizontal subproject(s) with a technical content focussed
on original research work of common interest in these fields, leading to
tangible results.

These six months were a period of intensive work, which included the organization of three workshops regrouping partners coming from industry and public research. ProFusion first phase was a success (congratulation from the PReVENT core group). The ProFusion second phase should start at the beginning of 2005.

Puvame [October 2003-September 2005]

National project, Predit Programme Puvame ``Protection des Usagers Vulnérables par alarme et Manoeuvre d'Evitement''

An important number of accidents between vulnerable road users and moving traffic could be avoided by improving the abilities of visibility and estimation of the situation by the driver, and by putting in action an alarm system addressed to the driver and the road user in danger. This project will contribute to reduce the number of accidents of this type, by developing the principal following functionalities:
(1) Improvement of the abilities of perception of the driver in close and average distance environments by dated fusion;
(2) Detection and estimation of the dangerous situations, by analyzing current data relating on the "behavior of the driver" and to the estimation result of the "dangerosity" of the operations in progress;
(3) Activation of alert actions associated to vehicle and vulnerable users;
(4) Integration and experiments on vehicles and preliminary study on bus and/or trams.
INRIA Rhône-Alpes is coordinator of the project. The partners are:
*e-Motion* project (Inria Rhône-Alpes; Christian Laugier, Olivier Aycard, Thierry Fraichard, Anne Spalanzani)
and Imara project (Inria Rocquencourt; Michel Parent), Ecole des Mines
de Paris (EMP), INRETS, Intempora, Probayes, Robosoft, Connex.

Mobivip [October 2003-September 2005]

National project, Predit Programme Mobivip ``Véhicules Individuels Publics pour la Mobilité en centre ville''

The project gathers 5 laboratories and 7 industrials to implement, evaluate and demonstrate the NTIC impact on a new mobility service. More precisely, the goals are to implement:(1) a transportation service base on free-use vehicles, (2) a multimodal information system, (3) a toolbox for integration in global management policy at downtown scale.

ARCOS [June 2003-December 2004]

National project, Predit Programme Arcos``Action de Recherche pour une COnduite Sécurisée''

Within the PREDIT programme, the ARCOS project has for aim to find a global solution for the system `` vehicle-driver-infrastructure' ', in order to contribute to the improvement of the road safety(with the goal to reduce by 30% the accidents!). *e-Motion* is in charge of the research subject `` Information synthesis and commands development''. This project includes many french laboratories working in the field of vehicles and road (ENSMP, INRETS/LIVIC, SUPELEC, UTC...) as well as the largest french car manufacturers (PSA and Renault).

Kelkoo [October 2003-March 2004]

Industrial project, ProBayes start-up.

Kelkoo the european leader of price comparison on Internet proposes to its customers a service making it possible to compare the offers on the market. In order to propose increasingly precise information on the products Kelkoo creates a data base of these products. A very important function is the association of offers with the products of Kelkoo data base. This function already exists but requires many adjustments to take into account the inaccuracies of the different offers suggested. The aim of the collaboration of Kelkoo with INRIA is the validation of the use of an inference engine based on bayesian probabilities to carry out a matcher able "to learn" and to improve the results of the matching offer/product without requiring long and expensive adjustments.

AMIB-E [June 2002-June 2004]

This RNTL collaboration project (called AMIB-E) between INRIA and two industrial partners,(n amely: MGE UPS SYSTEMS and TEAMLOG) aims at building a Bayesian preventive maintenance tool.

The first prototype concerns the modeling and the simulation of the wearing process of the UPS capacitance component.Its aim is to warn the system maintainer that this component have to be replaced when the probability to get this component out of order(after a given period time) reaches a given threshold value.The next phase in this project will concern modeling more complex UPSsystems. To make this modeling task easier, a user-friendly graphic interface will be developed. Its aim is to allow graphical specification of the probabilistic model followed by an automatic generation of the corresponding ProBT code.

Visteo [May 2000-July 2004]

This project has started in May 2000 for a 24 months-length. This project is
supported by the PRIAMM national programme. The intial partners of the
project was GETRIS images, the Sharp project (now *e-Motion*) and the Movi
project of INRIA Rhône-Alpes. After one year of work, the project has been
``frozen'' until the GETRIS images company has been replaced in 2003 by a
new one: XL-Studio located in Lyon, with an extension of the project for 18
months. The aim of this project is to develop a set of software tools
allowing the set up of ``virtual studios'' physically realistic and
including interactions between the virtual character and the human
displayed in the studio. The project was succefully terminated at the end of July 2004, with a transfert of the *VisteoPhysic* library to the XL-Studio company.

Robea ``Speech-gifted android'' [2002-2004]

This project gathers three partners : ICP-IMAG, University of Texas At Austin, and Inria *e-Motion*. Its objectives are the study and modelling of perception,
production and learning mechanisms, in order to improve our understanding
of the word and the language, and to open new ways for their automatic
treatment.

Robea ``Bayesian models for motion generation'' [2003-2005]

This project involves three partners : Inria *e-Motion* (leader),
Inria *Evasion*, and Irisa *Siames*. The objective is to propose new forms of interaction between human and data-processing systems. The synthetic worlds created and managed by these systems can be populated by human actors and virtual actors controlled by
computers. The approach that we propose consists in equipping the virtual entities in these environments with an autonomy of movement and action, as well as adaptability and reactive abilities to certain situations.

Robea ``Parknav'' [October 2002- September 2005]

This project gathers five partners : Inria *e-Motion* (leader),
Inria *Movi* and *Prima*, Irisa *Vista*, and LAAS Toulouse.
The goal of the project is to automate the driving of a vehicle
moving amidst mobile obstacles (other vehicles, pedestrians) on a site
equipped with a camera-based perception system. A joint demonstration integrating both the perception and the planning levels is scheduled in the last phase of the project. It will take place on the Inria Rhône-Alpes parking site which is currently being equipped with a
multiple-camera perception system and will involve Cycab experimental vehicles.

IST-FET ``Biba'' [November 2001- November 2005]

European project IST-2001-32115 ``Bayesian Inspired Brain and Artefacts''.
The project involves five partners : Inria *e-Motion* (coordinator), University College
of London (Gatsby Unit), University College of Cambridge (Physiology lab),
Collège de France (Laboratory of Physiology of Perception and Action, A. Berthoz),
Ecole Polytechnique Fédérale de Lausanne (Autonomous Systems Lab, R. Siegwart), and the
Massachusetts Institute of Technology (Non Linear Systems Lab, JJ. Slotine).
The twin technological and scientific goals of the *BIBA* project are the followings :(1) to reconsider in the light of Bayesian probabilistic reasoning our
methodology, models,algorithms and techniques for building artefacts for
the ``real world''; (2) to provide a firm Bayesian basis for understanding
how biological systems use probabilistic logic to exploit the statistical
properties of their environments. The project is organised along 3 axes of research and development: (1) Neural basis of probabilistic inference; (2) New probabilistic models and
algorithms for perception and action; (3) New probabilistic methodology and
techniques for artefact conception and development.

France-Mexico ``Navdyn'' [October 2003- October 2005]

The *NavDyn* project is a joint *Lafmi**e-Motion* and the Center for Intelligent System (CSI) of the Mexican Technological Institute of Monterrey (ITESM). The goal of the project is to develop basic tehnologies for the ``Autonomous Navigation in Dynamic Environments''. CSI Itesm is in charge of the vision part of the project (detection and tracking of moving objects using an off-board pan-tilt video camera), whereas *e-Motion* is in charge of the autonomous motion part (taking into account moving objects with unknown future behaviour). The midterm evaluation that took place in November 2003
was successful and the project was prolongated.

France-Singapore ``Neurophysiology and robotics'' [June 2004- June 2008]

This CNRS-PICS project involving the Collège de France (LPPA), INRIA *e-Motion*, the University of Singapore (NUS), and the *IPAL* joint CNRS-NUS laboratory in Singapore. The objective is to study some aspects of the physiology of human vision, and to try to develop robotics models inspired from biological systems. An application of this research we will be to control a wheelchair from natural human control channels. This research involves a co-directed PhD student located in singapore.

France-Korea ``SafeMove'' [April 2004- April 2006]

The *SafeMove* project is a joint project in the scope
of the France-Korea STAR programme. It aims at conducting common research
activities in the area of Intelligent Transportation Systems (ITS) and
Automated Guided Vehicles (AGV). The proposed project combines three research groups (two French: Inria and
Lasmea Clermont-Ferrand, and one Korean: Sungkyunkwan University) having
complementary skills and expertise to conduct research in the area of ITS
and AGV, particularly focused on models and algorithms allowing for safe
autonomous navigation in dynamic environments (like those found in a urban
context).

Since October 1997, *e-Motion* has a collaboration with Riken Institute in Tokyo in the multi-robots systems field . Crossed visits have occured for the last three years. In 2002, Thierry Fraichard spent 4 months in Riken Institute and Igor Paromtchik spent 4 months at INRIA. Informal exchanges are still going on.

The common laboratory, named *Intelligent Vehicle Lab*, between
Nanyang Technological University of Singapore (NTU) and Inria has
started in November 1998, in the framework of the scientific
collaboration in the field of autonomous vehicles. This collaboration
has brought: (a) an important number of crossed visits and stays (one
week to several months) of researchers, (b) Singaporeans students in Inria (level undergraduate to graduate), and (c) organization of workshops (1999-2002). in 2003, Julien Diard has been a Postdoc student in NUS and a co-directed PhD (Brice Rebsamen) will begin in January 2004 in NUS. In addition, a PICS CNRS project has been accepted in 2004 with NUS and LPPA(College de France, Alain Berthoz).

Collaboration in the field of dextral handling begun with the stay of professor K. Gupta at Inria Rhone-Alpes in 1995, continued by several long stays of Moez Cherif and Juan Manuel Ahuactzin, and with crossed visits. Common publications has been done in 2000 and 2001. Informal exchanges are still going on.

In the framework of the programme France-Berkeley a one-year collaboration in the field of dynamic simulation has been settled in 2000-2001. This collaboration continued in 2002 within the framework of a new project on the simulators of surgery supported by the France-Berkeley funds (with the team S. Sastry). Several crossed stays of researchers took place. David Bellot is now a Postdoc student in Berkeley.

The study of force-feedback in virtual environments and the non linear elastic deformations have been the research subjects of this collaboration with the Center of Advanced Technology in Surgery of Stanford University. Several researchers crossed stays also took place (in particular: Remis Balaniuk).

The thematic network "Image et Robotique" has been implemented from the French-Mexican symposium in Computer Sciences and Control (JFMIA'99) which has been held in Mexico in March 1999. The main goal of this network is to promote and increase the french-mexican cooperations in Image and Robotics in scientific, academic and industrial fields. This network has been effectively settled in 2000. It supports a yearly school, students exchange and crossed visits.

The NavDyn project between *e-Motion* and ``Centro de Sistemas
Inteligente'', Itesm, Monterrey lasts from October 2002 to September 2004. This project supported by the French-Mexican Lab in Computer Sciences (LAFMI) studies the field of vehicle navigation in dynamic environment.

The dissemination of results and the active participation to international
scientific events (see bibliography) are two essential activities of the
*e-Motion* project-team.

Some members of *e-Motion* participates to some international committees:

C. Laugier is a member of the steering-advisory committee of IEEE/RSJ IROS (Intelligent Robots and Systems) international conference since 1997. He is also a member of the advisory committee of the ICARCV International conference on Control, Automation, Robotics and Vision.

C. Laugier is a member of the steering committee of the European Network EURON.

C. Laugier is a member of the following national scientific committees : French-Korean committee of the French Ministry of Foreign Affairs, National programme in Robotics ROBEA, CNRS RTP17 on Robotics, and inter-ministerial PREDIT group 9 on new technologies for transport.

T. Fraichard was member of the program committee of the Iberoamerican Conf. on Artificial Intelligence, Puebla (MX), November 2004, and the IEEE-RSJ Int. Conf. on Intelligent Robots and Systems. Sendai (JP), October 2004.

T. Fraichard was a member of the organization committes of the First Korea-France Symposium on Dependable Robotic Navigation, Seoul (KR), October 2004.

T. Fraichard and P. Bessière are regurlarly members of the programme committees of the ICRA and IROS conferences.

In addition to ponctual academic lectures, the members of *e-Motion* have tought the following lectures:

Lecture ``Motion planning'', LAFMI Summer School on Image and Robotics: Sophia
Antipolis (FR) [June 2004]. *Teacher: T. Fraichard.*

Lecture ``Robotic''s, Summer school "Automatic Control for Production Systems",
Grenoble [June 2001, July 2002, June 2004]. *Teacher: T. Fraichard.*

Lecture ``Motion planning'', Master "Mathématics, Informatics" , "Image, Vision,
Robotics" option (ex DEA), INPG-UJF, Grenoble. *Teacher: T. Fraichard.*

Lecture ``Outils probabilistes et statistiques pour l'Informatique'': (every year): Licence d'Informatique de l'Université Joseph Fourier 3^{e} année, Grenoble, (FR). *Teacher: O. Aycard*.

Lecture ``Inférence et apprentissage bayesien'': (every year): Ecole Polytechnique de Grenoble, Grenoble, (FR).
*Teacher: O. Aycard*.

Lecture ``Inférence et apprentissage bayesien": (every year): Master d'Informatique de l'Université
Joseph Fourier 1^{re} année, Grenoble, (FR). *Teacher: O. Aycard*.

Lecture ``Bayesian robot programming'': (every year): Master 2^{e} année ``Imagerie,
Vision, Robotique'' de l'INPG, Grenoble, (FR). *Teachers: P. Bessière and O. Aycard*.

Lecture ``Bayesian robot programming'': France-Mexico Summer school on
``Image and Robotics'' (every year). *Teachers: O. Aycard and P. Bessière*.

Tutorial on ``Autonomous mobile robots in real environments'': Singapore, May 2004. *Teachers: C. Laugier and O. Aycard*.

Lecture ``Introduction to robotics and current research issues'' (every year): France-Mexico Summer school on ``Image and Robotics'' (every year). *Teacher: C. Laugier.*

Lecture ``Robotics and motion autonomy'' (every year): DEA ``Imagerie, Vision,
Robotique'' INPG, Grenoble, (FR). *Teacher: C. Laugier*.

Lecture ``Basic tools and models for Robotics'' (every year): Cnam Grenoble. *Teachers: C. Laugier and J. Troccaz.*

Some members of *e-Motion* participates to various international conferences committees and to the organization of summer schools :

C. Laugier and J.M. Ahuactzin participated to the organizing committee of the 2004 summer school on ``Image and Robotics'' at Inria Sophia-Antipolis (June 2004).

C. Laugier participate every year to the organization committees of the major international conference on Robotics, in particular : IEEE International Conference on Robotics and Automation (ICRA), IEEE/RSJ Intenational Conference on Intelligent Robots and Systems (IROS), International Conference on Field and Service Robotics (FSR). He was general chair of IROS'97, Regional programm chair of IROS'00, Programme chair of IROS'02.

T. Fraichard and P. Bessière are regurlarly members of the programme committees of the IEEE ICRA and IEEE/RSJ IROS conferences.

Invited session organisation, Autonomous vehicles in dynamic environments, Int. Conf. on Control, Automation, Robotics and Vision, Kunming (CN), December 2004, by Th. Fraichard.

Some members of the *e-Motion* project-team have also given invited talks adressed to a larger public :

C. Laugier gave a talk at the conference SRG'04 in Singapore (May 2004) on ``Will future robots really share our living space ?''.

Seminar ``TIC et mobilité'', Antibes (June 2004). Talk on ``Tools for automated driving'' by Th. Fraichard.

Colloquium of the French Navigation Institute, Paris (March 2004). Talk on ``Intelligent transport systems'' by Th. Fraichard.