Probabilistic Reasoning and Decision Making in Sensory-Motor Systems

We assume that both living creatures and robots must face the same fundamental difficulty: incompleteness (and its direct consequence uncertainty).

Any model of a real phenomenon is incomplete: there are always some hidden variables, not taken into account in the model, that influence the phenomenon. The effect of these hidden variables is that the model and the phenomenon never behave exactly alike. Both living organisms and robotic systems must face this central difficulty: how to use an incomplete model of their environment to perceive, infer, decide and act efficiently.

The purpose of this chapter is to introduce gently the basic concepts of Bayesian programming.

We also present the technical issues related to Bayesian programming: inference principles and algorithms and programming language. Although they are very interesting, we have kept this part very short, as these technical questions are not central to this book.

Autonomous navigation of a mobile robot is a widely studied problem in the robotics community. Most robots designed for this task are equipped with onboard sensor(s) to perceive the external world (sonars, laser telemeters, camera). Two main approaches to autonomous navigation have been proposed: reactive navigation, where the robot uses only its current perceptions to move and explore without colliding (e.g. Arkin (1998) or Bonasso et al. (1995)), and servoed navigation, in which the robot is given a preplanned reference trajectory and uses some closed-loop control law to follow it (e.g. Laumond et al. (1989) or Lamiraux et al. (1999)). In servoed problems, two classes of approaches can again be separated: state-space tracking (e.g. Hermosillo et al. (2003b,a)) and perception- space tracking (e.g. Malis et al. (2001) or Chaumette (1994)).

Perception of and reasoning about dynamic environments is pertinent for mobile robotics and still constitutes one of the major challenges. To work in these environments, the mobile robot must perceive the environment with sensors; measurements are uncertain and normally treated within the estimation framework. Such an approach enables the mobile robot to model the dynamic environment and follow the evolution of its environment. With an internal representation of the environment, the robot is thus able to perform reasoning and make predictions to accomplish its tasks successfully. Systems for tracking the evolution of the environment have traditionally been a major component in robotics. Industries are now beginning to express interest in such technologies. One particular example is the application within the automotive industry for adaptive cruise control (Coué et al., 2002), where the challenge is to reduce road accidents by using better collision detection systems. The major requirement of such a system is a robust tracking system. Most of the existing target-tracking algorithms use an object-based representation of the environment. However, these existing techniques must explicitly consider data association and occlusion. In view of these problems, a grid-based framework, the Bayesian occupancy filter (BOF) (Couéet al., 2002, 2003), has been proposed.

Mobile robots are gradually appearing in our daily environment. To navigate autonomously in real-world environments and interact with objects and humans, robots face various major technological challenges. Among the required key competencies of such robots is their ability to perceive the environment and reason about it, to plan appropriate actions. However, sensory information perceived from real-world situations is error prone and incomplete and thus often results in ambiguous interpretations. We propose a new approach for object recognition that incorporates visual and range information with spatial arrangement between objects (context information). It is based on using Bayesian networks to fuse and infer information from different data.In the proposed framework, we first extract potential objects from the scene image using simple features- characteristics like colour or the relation between height and width. This basic information is easy to extract but often results in ambiguous situations between similar objects. To resolve ambiguities among the detected objects, the relative spatial arrangement (context information) of the objects is used in a second step. Consider, for example, a cola can and a red trash can that are both cylindrical, have similar ratios between width and height and have very similar colours. Depending on their distances from the robot, they may be hard to distinguish. However, if we further consider their spatial arrangement with other objects, e.g. a table, they might be clearly differentiable, the cola can typically standing on the table and the trash can on the floor. This contextual information is therefore a very efficient way to increase drastically the reliability of object recognition and scene interpretation. Moreover, range information from a laser scanner and speech recognition offer complementary information to improve reliability further. Thus, an approach using laser range data to recognize places (such as corridors, crossings, rooms and doors) using Bayesian programming is also developed for both topological navigation in a typical indoor environment and object recognition.

Imagine yourself lying in your bed at night. Now try to answer these questions: Is your body parallel or not to the sofa that is two rooms away from your bedroom? What is the distance between your bed and the sofa? Except for some special cases (like rotating beds, people who actually sleep on their sofas, or tiny apartments), these questions are usually nontrivial, and answering them requires abstract thought. If pressed to answer quickly, so as to forbid the use of abstract geometry learned in high school, the reader will very probably give wrong answers.

However, if people had the same representations of their environment that roboticians usually provide to their robots, answering these questions would be very easy. The answers would come quickly, and they would certainly be correct. Indeed, robotic representations of space are usually based on large-scale, accurate, metric Cartesian maps. This enables judgment of parallelism and estimations of distances to be straightforward.

What similarities can be found between an animal and an autonomous mobile robot? Both can control their motor capabilities based on information acquired through dedicated channels. For an animal, motor capabilities are muscles and joints, and filtered information from the environment is acquired through sensors: eyes, nose, ears, skin, and several others. For a mobile robot, motor capabilities are mostly end effectors and mechanical motors, and information about the surroundings consists of data coming from sensors such as proximeters, laser range sensors and bumpers.

We present “BCAD”, a Bayesian CAD modeller for geometric problem definition and resolution. This modeller provides tools for (i) modelling geometric uncertainties and constraints, and (ii) solving inverse geometric problems while taking into account the propagation of these uncertainties. The proposed method may be seen as a generalization of constraint-based approaches in which we explicitly model geometric uncertainties. Using our methodology, a geometric constraint is expressed as a probability distribution on the system parameters and the sensor measurements instead of as a simple equality or inequality. To solve geometric problems in this framework, we propose the Monte Carlo Simultaneous Estimation and Maximization (MCSEM) algorithm as a resolution technique able to adapt to problem complexity. Using three examples, we show how to apply our approach using the BCAD system.

This work is within the context of Computer-assisted Orthopaedic Surgery (CAOS), in particular Total Hip Replacement (THR) surgery.

“Computer-assisted Surgery (CAS) has the aim of assisting surgeons in their therapeutic efforts to be as exact and minimally invasive as possible”(Corbillon, 2002). CAS is an interdisciplinary research area; it uses many sources of information, devices, computer techniques and clinics. In the past, medical imagery was only used for diagnosis and pathology localization. Today, image processing and computer-assisted surgery systems help surgeons to improve their perception and action capabilities.“Medical image processing makes possible the acquisition of a numerical model of reality. In surgery, this corresponds to a replica of the patient’s anatomy”(Corbillon, 2002).

Today’s video games feature synthetic characters involved in complex interactions with human players. A synthetic character may have one of many different roles: a tactical enemy, a partner for the human player, a strategic opponent, a simple unit among many, or a substitute for the player when he or she is unavailable.

In all of these cases, the game developer’s ultimate objective is for the synthetic character to act as if it were controlled by a human player. This implies the illusion of spatial reasoning, memory, commonsense reasoning, using goals, tactics, planning, communication and coordination, adaptation, unpredictability, and so on. In current commercial games, basic gesture and motion behaviours are generally satisfactory. More complex behaviours usually look much less lifelike. Sequencing elementary behaviours is an especially difficult problem, as compromises must be made between too-systematic behaviour that looks automatic and too-random behaviour that looks ridiculous.

In addition to the five senses usually described, vertebrate species possess a sensory organ that detects motion of the head. This organ is the vestibular system, located in the inner ear. Motion information collected by the vestibular system is crucial for equilibrium. It also contributes to stabilizing the gaze in space during head movements. Motion information provided by the vestibular system generates compensatory eye movement, a phenomenon called the Vestibulo-Ocular Reflex (VOR). The importance of this function is illustrated by the following example (from Guedry (1974)): you can look at the lines on your hand and shake your head at the same time. The VOR provides efficient gaze stabilization in this condition. In contrast, if you shake your hand, looking at the lines becomes impossible.

We use multiple sensory modalities to perceive our environment. One of these is optic flow, the displacement and deformation of the image on the retina. It is generally caused by a relative motion between an observer and the objects in the visual scene. As optic flow depends largely on three-dimensional (3D ) shapes and motions, it can be used to extract structure from motion (the sfm problem). Motion parallax and the kinetic depth effect are special cases of this phenomenon, noticed by Von Helmholtz (1867), and experimentally quantified by Wallach and O’Connell (1953).

Speech is a perceptuo-motor system. A natural computational modelling framework is provided by cognitive robotics, or more precisely speech robotics, which is also based on embodiment, multimodality, development, and interaction. This chapter describes the bases of a virtual baby robot, an articulatory model that integrates the non-uniform growth of the vocal tract, a set of sensors, and a learning model. The articulatory model delivers sagittal contour, lip shape and acoustic formants from seven input parameters that characterize the configurations of the jaw, the tongue, the lips and the larynx. To simulate the growth of the vocal tract from birth to adulthood, a process modifies the longitudinal dimension of the vocal tract shape as a function of age. The auditory system of the robot comprises a “phasic” system for event detection over time, and a “tonic” system to track formants. The model of visual perception specifies the basic lip characteristics: height, width, area and protrusion. The orosensorial channel, which provides tactile sensations on the lips, the tongue and the palate, is elaborated as a model for the prediction of tongue–palatal contacts from articulatory commands. Learning involves Bayesian programming, in which there are two phases: (i) specification of the variables, decomposition of the joint distribution and identification of the free parameters through exploration of a learning set; and (ii) utilization, which relies on questions about the joint distribution.