Professional Documents
Culture Documents
1993 Simulating Physical Interactions Between An Articulated Mobile Vehicle and A Terrain
1993 Simulating Physical Interactions Between An Articulated Mobile Vehicle and A Terrain
Elsevier
Abstract
Jimenez, S., Lueiani, A. and Laugier, C., Simulating physical interactions between an articulated mobile vehicle and a
terrain, Robotics and Autonomous Systems, 11 (1993) 97-107.
This paper deals with the problem of planning the motion of a complex land vehicle moving in a natural environment. The
contribution presented here is a motion generator which predicts the dynamic behaviour of the vehicle when executing a
given nominal motion plan. This plan is expressed in terms of a channel that must be followed and of a set of intermediate
subgoals that must be reached. Solving this motion generation problem requires to explicitly reason about the geometric and
the physical aspects of the movements that the vehicle has to execute. In our approach, this is done by using two basic
constructions derived from the concept of physical model: the 'generalized obstacles' are used to physically guide the
movements of the vehicle using an explicit model of the vehicle-terrain interactions, and the 'physical targets' are used to
map the strategic information onto our physical representation of the world and to introduce the force feedback gestural
control.
Keywords: Motion generator; Predictive force-display; Physical models; Generalized obstacles; Physical targets;
Vehicle/terrain physical interactions; Force feedback gestural control
this project is to design a complete system for in this context - because friction, sliding and skid
teleprogramming an autonomous planetary rover. phenomena may modify the behaviour of the
The contribution presented in this paper is a vehicle -, it is necessary to make use of appropri-
motion generator which predicts the dynamic be- ate physical models for predicting the robot be-
haviour of a land vehicle when executing a given haviour.
nominal motion plan. This plan is supposed to be This is why we have decided to combine a
provided by a geometric path planner. It is ex- geometric path planner with a motion generator
pressed in terms of a channel that must be fol- which has the ability to simulate physical interac-
lowed and of a set of intermediate subgoals that tions. This approach is described in Section 2. It
must be reached. Solving this motion generation basically consists of combining strategic and tacti-
problem requires to explicitly reason about the cal information through two complementary con-
geometric and the physical aspects of the move- cepts (see Section 4); the concept of generalized
ments that the vehicle has to execute. obstacle which is used for physically guiding the
Until now, very few methods have been de- vehicle using an explicit model of the vehicle-ter-
vised for dealing with the effects of dynamics in rain interactions, and the concept of physical
motion planning. Some researchers have de- target which allows us to map the strategic infor-
signed adapted architectures for controlling in a mation (intermediate subgoals) onto our physical
dynamic way the locomotion of several types of representation of the world. These concepts have
mobile robots (legged robots for instance [13]) been implemented using physical based models
without trying to solve the associated motion- first developed in the field of computer graphics
planning problem. Because of the intrinsic com- and of man-machine communication [8,11]. As
plexity of this problem most of the work done on shown in Section 3, such models can be built by
this topic has led to the implementation of vari- combining a set of representative punctual masses
ous geometric path planners (the considered sub- within an interaction network including linear
problem is to find a collision-free trajectory for and non-linear terms. In Section 5 we explain
the robot) [9]. More recent methods have been how this new approach can lead toward a power-
devised for the purpose of dealing with non-inte- ful extension of the 'Predictive Display' and to a
grable kinematic constraints when generating such realistic reconstruction of the natural piloting
paths (see [1,6,10]). Despite the great ability of conditions by means of adapted force feedback
the previous planners to solve such geometric devices. Several types of vehicle and of vehicle-
based problems, the automatic generation of exe- terrain interactions have been modelized and
cutable safe motions for a land vehicle moving in simulated using this approach (see Section 6).
a natural environment is beyond the current ca-
pabilities of these planners (because the physical
vehicle-terrain interactions cannot be processed 2. Outline of the approach
using purely geometric and kinematic models).
Since such interactions are of major importance As previously mentioned, our approach for
planning executable safe motions for a land vehi-
cle moving in a natural environment is to com-
Christian Laugier is Research Direc- bine a geometric path planner with a motion
tor at INRIA (Institut National de
Recherche en Informatique et Au- generator which has the ability to simulate physi-
tomatique) and director of the cal interactions. The geometric path planner is in
robotics group at L I F I A / I N R I A
Rh6ne-Alpes. He was associate direc- charge of generating a potentially safe path for
tor of the L I F I A laboratory the rover, i.e. a set of collision-free trajectories.
(Laboratoire d'Informatique Fonda-
mentale et d'Intelligence Artificielle) Such a path is represented by a safe channel C
from May 1987 to December 1991. and by a set (gl, g2.-. gn) of intermediate sub-
He also participated to the start-up of
three industrial companies in 1982, goals. It defines strategic information that the
1985, and 1989 respectively. His cur- motion generator will use to produce the se-
rent research interests are broadly in the areas of automatic
robot programming, autonomous robot vehicles, and motion quence of movements which results from the
planning. Laugier received the MS, Ph.D. and Docteur d'Etat combined effects of the applied motion com-
degrees in computer science from Grenoble University in
1973, 1976, and 1987 respectively. mands and of the vehicle-terrain interactions
Physicalsimulationof planetary mobilevehicles 99
a) b)
Fig. 1. Illustrationof our approach: (a) a selected safe channel; (b) an associated physicalrepresentation; (c) a physicaltarget
attracting the vehicle;(d) a motionresultingfromthe combinedeffectsof the physicaltarget and of the vehicle-terrain interaction.
(see Fig. 1). This approach which consists in mechanisms have to be developed in order to
combining a global path planner with a method map the geometric information contained in SP
aimed at adapting the robot movements to the onto ~(M, T). As explained in Section 4, this is
local environment of the task has already been done using two basic constructions: q~(T) and
used by some researchers for implementing geo- • (C) are combined within a unified model qb(T
metric path planners and reactive motion con- C) called the 'generalized obstacle model', and
trollers (see [7]). But the models used - mainly each gi is converted into an attractive punctual
geometric representations and potential field mass ~(gi) called 'physical target'. Using this
functions - do not include the basic information approach, ~ ( T ~ C) can be used for physically
which is required for processing physical interac- guiding M, and the set {~(gt), ~(gE)..-qb(gn)}
tions. can be used for representating the sequence of
Let M be the land vehicle, T the terrain, and motion commands to apply to M. This approach
O the obstacles to be avoided, and let SP= differs from the potential field methods by the
(C, g~, g2... gn) be the, potentially safe path gen- fact that it operates at a dynamic layer, i.e. any
erated by the geometric path planner. We will motion is directly connected by physical laws to
respectively denote by G(W) and by qo(W) the the forces which result from object interactions
geometric model and the physical model which is (see Section 3).
associated to an object (or a set of objects) W. Since no complete connection has been imple-
The geometric path planner reasons about the mented yet between the geometric path planner
model (M, T, 0) and the motion generator oper- and the motion generator, the models ~ ( T • C)
ates on the model ~ ( M T, SP). This means that and {~(gt), ~(g2)...q~(gn)} are currently con-
G(M) and G(T) have to be converted into a structed by the operator using the modeling capa-
physical representation ~(M, T) including an ex- bilities of the Cordis-Anima system [8,11]. In the
plicit model of the vehicle-terrain interactions sequel, we will make the assumption that the data
(see Section 3). This also means that appropriate G(M), G(T) and SP are known beforehand, and
100 S. Jimenez et al.
C) "
a) ~ ~ ~ : : : : ~ . ~ Virtual
Real . ~ ~ Obstacles
Obstacle~ .---3
b, ~ ~ / D
~ Z , C ~, ~ r~ ~ Interaction
~ Borderline
i ~ ° ."~.]... Punctual
B ~ Mass
Fig. 2. (a) Numerical model and physical model of T. (b) Modeling an articulated mechanism. (c) A 'generalized obstacle '.
we will focus on the physical aspects of the mo- structural modifications (plasticity properties)
tion generation problem. [8,11].
In our model, any object is represented by a
structured set of spherical components (balls),
3. World modeling using physical models and any pair of interacting components is con-
nected by an appropriate connector. This ap-
3.1. Physical models proach allows us to process physical interactions
using some elementary dynamic equations involv-
Geometric models have been designed for rep- ing punctual masses only: a ball is seen as a pair
resenting shape properties and for defining (m, r), where m is an 'active' punctual mass
macroscopic behaviours for solid objects. But, located at the center of the ball and r is the
such models do not include the ability of process- radius of the associated interaction area (this
ing complex physical interactions, because the area is generally characterized by a non-penetra-
associated effects result from the integration of tion law) [8]. Consequently, the physical model
differential equations involving distributed force qb(W) of a set W of interacting objects can be
and position parameters. This is why researchers represented by a network in which each node
in the field of computer graphics and of man- n i ~ .A~ defines an object component (a ball), and
machine communications have developed the each arc aij ~ ' represents a physical interac-
concepts of 'physical model'. The basic idea is to tion. A node n i is characterized by a pair (Pi, m~),
consider that any object motion a n d / o r deforma- where p~ and mi are respectively the position and
tion is linked by physical laws to a set of forces the weight of the associated punctual mass; an
whose application points depend on the intrinsic arc aij defines a particular interaction equation
structure of the involved objects. A practical way of the type Fi = - F j = CI)ij(Pi, pj), where qbij rep-
for modelizing this phenomenon is to appropri- resents a linear damper/.spring law or a particu-
ately discretize the object and to characterize lar nonlinear interaction law.
each obtained component by differential equa-
3.2. Terrain and vehicle modeling
tions combining two dual variables: the force f
and the position P (or the velocity V). Such a As explained in Section 2, the geometric mod-
model is usually called 'continuous lumped con- els G(T) and G ( M ) have to be converted into a
stant model'. It basically permits the construction physical representation ~(T, M). The first step
of complex objects by assembling object compo- of the algorithm is to construct ~ ( T ) and ~ ( M )
nents using various connectors, for instance: con- using appropriate connected sets of balls.
nectors representing the non-linearities associ- Let G(T) be a numerical model of the surface
ated to the object boundaries (for modeling con- of T. Converting G(T) into q~(T) consists in
tact behaviours), or connectors defining some finding an appropriate agglomerate of balls in
Physical simulation of planetary mobile vehicles 101
which each ball represents a punctual mass and (2) Compute the new position Pi to associate to
its associated non-penetrating area - this area is each node n i in ~ ( W ) using the Newtonian
characterized by an elastic/viscous repulsive law. law:
In our system, this agglomerate is constructed
using a dynamic ball meshing algorithm which
consists in 'throwing' a set of interactive balls ~.t l ~ = m i . ~ i = m i . d / 2 -
Vaik
toward a rigid net representing the surface of T
[12] (see Fig. 2a). The size of the balls which are The same algorithm can be used for generating
used in this algorithm depends on the spatial the motions of the vehicle M by associating a
bandwidth of T and on the characteristics of M propeling force F to M. Since M has to take into
(the size of the wheels for instance). The interac- account to strategic information contained in SP
tion laws used to complete q~(T) are selected (see Section 2), F must be chosen in such a way
according to the physical characteristics of the that M will move toward the next subgoal gi
ground. Until now, most of the experiments have while following the channel C and reacting with
been processed using a rigid ground (see Section T. Two basic physical constructions have been
6). developed for mapping these motion constraints
G(M) includes shape and kinematics repre- onto ~(T, M):
sentations. This means that two main steps have
to be applied to construct ~(M): (1) selecting a The generalized obstacles
representative set of punctual masses, and (2) The first motion constraint to satisfy is to force
connecting theses punctual masses in order to M to follow C while reacting with T. This means
define rigid components and articulated mecha- that both ~(T) and q~(C) have to be used to
nisms. Rigid components are modelized using physically guide M. In other words, qb(T) and
almost indeformable spring connectors; articu- • (C) have to be combined within a unified model
lated mechanisms are constructed using appropri- q~(T ~ C) referred to as the 'generalized obstacle
ate damper/spring based structures (see Fig. 2b). model'. A practical way for implementing this
The last step of the physical modeling process capability is to consider that C is delimited by a
is to combine q~(T) and ~ ( M ) in order to explic- set {O1, 0 2 . . . Ore} of real or fictitious obstacles
itly represent the vehicle-terrain interactions. (see Fig. 1). Then, each 0 i can be converted into
This is done by adding new connectors between a model ~ ( 0 i) which can be easily combined with
some pairs of punctual masses belonging to ~(T) • (T): both ~ ( 0 i) and q~(T) are made up of
× ~(M). The interaction laws to associate to agglomerates of balls. The main difference be-
these connectors depend on both the physical tween • ( 0 i) and ~(T) lies in the fact that the
characteristics of the ground and the vehicle loco- interaction areas which are associated to the balls
motion mechanisms (wheeled vehicle, tracked ve- belonging to ~ ( 0 i) include a safety distance pre-
hicle, etc.). An implementation of a wheeled ve- venting M to collide with Oi (see Fig. 2c). Using
hicle moving on a rigid ground is described in this approach, it becomes possible to simultane-
Section 6. ously solve the locomotion problem and the ob-
stacle avoidance problem (see Section 6).
obtained in a similar way by combining the effects of physical targets for example, a genuine gestu-
of several physical targets. Thus, it becomes pos- ral relationship with the simulated environment
sible to constantly generate an appropriate pro- (see Fig. 3).
pelling force F for M by only evaluating the The function of the gestural transducer is to
current state of the target-vehicle interaction liaise the relation between the human operator
(see Section 6). Physical targets may also be inter- gesture and the numeric phenomena produced by
actively moved in the scene for the purpose of the computer. The gestural relationship is typi-
locally modifying the motion commands to apply cally bi-directional, and this kind of device must
to M. In the Cordis-Anima working environment, fulfil two main functions:
this can be done by using a real-time physical (1) to pick up the characteristic information about
simulator and a gestural retroactive device previ- the control gesture;
ously developed by the Acroe-Lifia Team (see (2) to set a mechanical resistance to the operator
Section 6). Such a capability is potentially useful gesture according to the numerical signal from
for interactively generating appropriate maneu- the simulation process. This last function is
vers in intricate situations - because force feed- fundamental since it provides the user with
back may help the operator - but it is not clear an interface allowing an efficient control of
(at the present time) how the resulting physical the simulated system.
data can be converted into a strategic geometric In 1978, the A C R O E research team has intro-
information for the planner. duced the idea of a gestural man-machine com-
munication with 'tactile feedback' in the frame-
work of sound synthesis, and more generally of
5. An extension of the 'Predictive Display' man-machine communication problems [3,4]. The
studies carried out during the past ten years have
The new approach we have presented in this led to the characterisation and design of the
paper can be replaced in the framework of an requisite device, called Modular Force Feedback
extension of the 'Predictive Display Tools'. The Transducer (MFFT). An international patent cov-
purpose is to extend the task performance visuali- ers both its modularity principle and its motor
sation functions by offering to the robotic opera- technology, which was especially developed to
tor a complete reconstruction of the natural pilot- achieve M F F T performances [2].
ing situation. This becomes feasible when using a Two typical situations constitute the frame-
motion generator based on dynamic models of work to determine the gestural device character-
the situation and connected with gestural devices istics. We want to have the ability to reinstall
for the tactile senses. good buffers in the places where collisions against
The whole system made of the modeler-simu- hard surfaces occurred. We also want to be able
lator for physical objects and of force feedback to restitute texture sensation in a general sense.
gestural devices allows, by means of manipulation Viscous and dry frictions are involved in these
[OPERA
•
OPERATOR
• ~ •
~
T
~"
"~
I
I
[ / R o b o t ~ T e&r r a i n
li.-I l ~ . ,.
i M
[]
I ~ ,,-, / Interaction H li
Displacement
Magnets sensor
15 mm
Coils
Magnetic field
Displacements
13.75 lnm
surface contacts. The sensor and especially the 13.75 mm and offers a o n e - D O F gestural chan-
motor performances can be deduced from these nel. The sixteen m o t o r - s e n s o r unit is organised
situations and from human gesture characteris- according to a linear placing of the sliced motors
tics. For example, the motor performance entails (Fig. 4 and Table 1).
an ultra-rapid, and accurate 'response' (in some The effective morphology is determined by the
cases the bandwidth of the mechanical phenom- association of a certain number of m o t o r - s e n s o r
ena concerned in the gesture can go as far as 700 modules from the sixteen-DOF unit and by the
or 800 Hz) and sizable power (up to pulse of 'Covering' that is given to the whole.
several tenths of KgF to simulate rigid obstacles). This 'Covering' consists of a very simple and
Moreover, it is not foreseeably possible to de- strongly built mechanical device chosen from an
sign a transducer for the gestural channel as 'assortment' of various possibilities which is
general as a loud-speaker is to the acoustic chan- mounted onto the m o t o r , s e n s o r module to com-
nel or a graphic screen to the visual channel. This bine.
remark therefore induced the A C R O E re- It is therefore possible to define a great num-
searchers to investigate what kind of device would ber of interchangeable covering like 1D, 2D and
guarantee a relative variety of gestures. The com- 3D covering. For example, in order to manipulate
plete device comes in the form of an extremely a 3D solid, we need a six-DOF device which will
compact unit built from sixteen sliced ' m o t o r - be obtained from the association of six m o t o r -
sensor modules'. Each slice has a thickness of sensor modules (Fig. 5).
Table 1
Motor-sensor module characteristics
Geometric Motor
mobile motor coil stroke 15 mm flat copper mobile coils vacuum soaked,
slice thickness 13.75 mm rare ground/cobalt magnets
~ 1D Stick
2D Stick
A pair of 2D Sticks: The Pliers
Fig. 5.1D and 2D joysticks. A mechanically connected transmission arm allows us to use one module as a 1D device or to combine
two independent motor-sensor modules to create a system with two degrees of freedom.
cation tools. Thus, complete information concern- staeles are represented by white circular regions.
ing the simulated physical process (animated Only the vehicle-obstacle interactions have been
graphic simulation, force feedback and states of considered in this experiment. These interactions
the dynamic parameters) can be used by the have been modelized using a classical repulsive
operator for evaluating the quality of the studied viscous-elastic law including some non-linearities
motion plan. (instantaneous changes between contact and
The geometric models of T and M and the non-contact states).
proposed motion plans are initially constructed The movements of the vehicle are generated
using our CAD-Robotics system (the ACT sys- here using an attractive physical target whose
tem) working on a Silicon Graphics workstation. position is interactively specified by the operator.
Figure 6 shows a visualization of a scene which The resulting motion makes the vehicle move
has been described using the geometric models of through the narrow corridor only if the target has
an articulated rover and of a Marsian landscape. been correctly positioned. The solution shown in
As previously mentioned, no complete connec- Fig. 7 has been easily found after a few attempts
tion has been implemented yet between the two performed in the simulated environment (i.e. in a
previous systems. Consequently, the required few seconds).
physical models are currently constructed by the
operator (within the Cordis-Anima environment) 6.3. Dynamic locomotion on a hilly terrain
from a description of the output of the ACT
system. Experiments involving several types of In this experiment, we have a 3D wheeled
vehicle (tracked vehicle, wheeled vehicle, articu- vehicle with several passive joints and six inde-
lated vehicle), several types of terrain, and sev- pendant conical wheels (see Fig. 8). The purpose
eral types of vehicle-terrain interactions have of the articulated mechanical structure is to pro-
been successfully processed. Some sequences of vide the rover with the capability of continuously
motions which have been generated by the system maintaining a sufficient number of contact points
for avoiding a set of obstacles and for moving a between the wheels and the ground when passing
rover on a solid hilly terrain are presented below. over some natural obstacles (rocks, small hills and
holes, etc.). This is achieved by automatically
6.2. Obstacle avoidance modifying the configuration of the articulated
mechanical structure using gravity and reaction
In this experiment, a train-like vehicle is mov- forces produced by the wheels-ground interac-
ing in a planar environment which contains a set tions.
of obstacles that have to be avoided (see Fig. 7). The terrain which has been used for this ex-
These obstacles are represented in the picture in periment has been modelized using an almost
grey color, and their associated generalized ob- rigid agglomerate of balls having non-penetrating
106 S. Jimenez et al.
properties. The vehicle has been modelized as required dynamic behaviour: a classical non-
explained in Section 3. Only the punctual masses penetrating model and a model aimed at generat-
which have been associated to the wheels are ing sliding-free rolling motions (this model com-
considered when computing the vehicle-terrain bines a friction term with the instantaneous tan-
interactions. Two basic interaction models have gential speed vector of the punctual mass which
been combined to provide the rover with the is involved in the current contact). The propelling
force which is applied to the rover is generated by tional project. It also has been partly supported
a physical target. Some steps of the resulting by the Rh6ne-Alpes Region through the
motion are shown in the Fig. 8. I M A G / I N R I A Robotics project SHARP.
7. Conclusion
References
We have presented a method for solving the
problem of planning the motions of a complex [1] J. Barraquand and J.C. Latombe, On non-holonomic
land vehicle moving in a natural environment mobile robots and optimal maneuvering, Revue d'Intelli-
gence Artificielle 3 (2) (1989) 77-103.
containing both obstacles that have to be avoided [2] C. Cadoz, L. Lisowski and J.L. Florens, Clavier retro-actif
and obstacles that can be passed over by the modulaire et actionneur modulaire extra-plat, French
vehicle. The main contribution of the paper is a Patent, Brevet INPI, Nb 88 14064 (13 Oct. 1988).
motion generator which predicts the dynamic be- [3] C. Cadoz, J.L Florens and A. Luciani, Responsive input
haviour of the vehicle when executing a given devices and sound synthesis by simulation of instrumental
mechanisms: the CORDIS system, CMJ 8 (3) (1984)
nominal notion plan. For that purpose, we have 60-73.
shown how the dynamic behaviour of such robots [4] C. Cadoz, L. Lisowski and J.L Florens, A modular
could be modelized and simulated using the con- feedback keyboard design, CMJ 4 (2) (1990) 47-51.
cept of 'physical model' which had initially been [5] Florens and C. Cadoz, The physical model. Modeling and
developed to solve computer graphics and man- simulating the instrumental universe, in: G. De Poli, A
Piciali and C. Roads, eds., Representation of Musical
machine communication problems. We have also Signals (MIT Press, Cambridge, MA, 1991) 227-268.
shown how strategic and tactical information con- [6] T. Fraichard, C. Laugier and G. Lievin, Robot motion
cerning the motion to be executed could be mod- planning: The case of non-holonomic mobiles in a dy-
elized and processed using this type of model. namic world, Proc. IROS'90, Tsuchiura, Japan (1990).
Several good experimental results obtained with [7] T. Fraichard and C. Laugier, On-line reactive planning
for a non-holonomic mobile in a dynamic world, Proc.
the Cordis-Anima modeler/simulator have fi- IEEE International Conference on Robotics and Automa-
nally been described. tion, Sacramento (1991).
The current work is oriented toward two main [8] S. Jimenez, Mod61isation et simulation d'objets volu-
directions: complementing the connection be- miques d6formables complexes (DEA technical report,
tween the path planner (strategic layer) and the INP-Grenoble, June 1989) (in French).
[9] C. Laugier, Reasoning in motion planning (Technical
motion generator (tactical layer), and developing report INRIA 932, Dec. 1988).
improved techniques for constructing a more [10] J.P. Laumond, M. Taix and P. Jacobs, A motion planner
complete terrain and other physical models for for car-like robots based on a mixed global/local ap-
vehicles. proach, Proc. IROS'90, Tsuchiura, Japan (1990).
[11] A. Luciani, S. Jimenez, O. Raoult, C. Cadoz and J.L.
Florens, A unified view of multiple behaviour, flexibility,
plasticity and fractures: Balls, bubbles and agglomerates,
Acknowledgement Proc. IFIP WG 5.10 International Conference on Modeling
in Computer Graphics (Springer Verlag, New York, (1991)
The research carried out by the A C R O E / 55-74.
LIFIA team is mainly supported by the French [12] J. Moutard, Planification et contr61e d'ex~cution simul~
d'un v6hicle autonome plan&aire (DEA technical report,
Minist~re de la Culture. The work presented in INP-Grenoble, June 1991) (in French).
this paper has been partly supported by the CNES [13] M. Raibert, Dynamic legged robots, Proc. ISER'88, Syd-
(French Spatial Agency) through the RISP na- ney, Australia (1988).