1

2
3
4
5
6 Automatica 38 (2002) 125}130
7
8 Brief Paper
9
10 A hybrid control approach to action coordination for mobile robots夽
11
12 Magnus Egerstedt , Xiaoming Hu
*
13
14 Division of Engineering and Applied Science, Harvard University, Cambridge, MA 02138, USA

Division of Optimization and Systems Theory, Royal Institute of Technology, SE-100 44 Stockholm, Sweden
15
16 Received 12 October 1998; revised 28 May 2001; received in "nal form 19 June 2001
17
18 Abstract
19
20 In this paper, the problem concerning how to coordinate the contributions from concurrent controllers, when controlling mobile
21 robots, is investigated. It is shown how a behavior based control system for autonomous robots can be modeled as a hybrid
22 automaton, where each node corresponds to a distinct robot behavior. This type of construction gives rise to chattering executions,
23 but it is shown how regularized automata can be used to solve this problem. As an illustration, the obstacle-negotiation problem is
24 solved by using a combination of a robust path-following behavior and a reactive obstacle-avoidance behavior that move the robot
25 around a given obstacle at a prede"ned safety distance.  2001 Published by Elsevier Science Ltd.
26
Keywords: Mobile robots; Hybrid dynamic systems; Motion control; Path planning
27
28
29
30 1. Introduction However, within this framework, a number of design 59
31 issues need to be addressed. Those range from questions 60
32 For mobile, autonomous robots the ability to function concerning the design of the individual behaviors to 61
33 in, and interact with, a dynamic, changing environment coordination issues. For instance, given a reactive ob- 62
34 is of key importance. A successful way of structuring stacle-avoidance behavior, modeled as a repulsive 63
35 the control system in order to deal with this problem potential "eld surrounding the obstacle, how should an 64
36 is within a behavior based control architecture. (See approach-target behavior be designed so that it takes 65
37 for example, Arkin, 1998; Kortenkamp, Bonasso, advantage of the fact that it is going to run in parallel 66
38 & Murphy, 1998). The main idea is to identify di!erent with an obstacle-avoidance behavior? Furthermore, how 67
39 controllers, responses to sensory inputs, with desired should these behaviors be combined? Addressing this last 68
40 robot behaviors. A behavior could, for instance, be question will be the main point of this paper. In other 69
41 obstacle-avoidance in which sonar information about words, via a case study where the robot is negotiating 70
42 a close obstacle should result in a movement away from obstacles, we will try to answer the question: How should 71
43 that obstacle. This way of structuring the control system the diwerent behaviors be combined so that the resulting 72
44 into separate behaviors, dedicated to performing certain robot response is satisfactory from both a safety and a 73
45 tasks has the major advantage that it makes the system performance perspective? 74
46 modular, which both simpli"es the design process as well 75
47 as o!ers a possibility to add new behaviors to the system 76
48 without causing any major increase in complexity, as 77
49 pointed out in Brooks (1986). 2. Path following and obstacle negotiation 78
50 79
51 夽
This paper was not presented at any IFAC meeting. This paper was
The speci"c problem that will be used to illustrate our 80
52 recommended for publication in revised form by Associate Editor Jurek point about action coordination is how to move a robot 81
53 Z. Sasiadek under the direction of Editor Mituhiko Araki. This work between two points in the plane. This point-to-point 82
54 was sponsored in part by the Swedish Foundation for Strategic motion should be done so that the detection of an ob- 83
55 Research through its Centre for Autonomous Systems at KTH and in stacle results in a repulsive potential "eld, acting on the 84
part by TFR.
56 * Corresponding author.
platform when the robot is closer to the obstacle than 85
57 a desired safety distance, d , where the subscript stands 86
E-mail addresses: [email protected] (M. Egerstedt), -
58 [email protected] (X. Hu). for obstacle-avoidance. This behavior is an example of 87
88
0005-1098/02/$ - see front matter  2001 Published by Elsevier Science Ltd. 89
PII: S 0 0 0 5 - 1 0 9 8 ( 0 1 ) 0 0 1 8 5 - 6 90
91
 126 M. Egerstedt, X. Hu / Automatica 38 (2002) 125}130

1 a so-called reactive obstacle-avoidance behavior. The 57

2 word reactive, a commonly used one in the robotics 58
3 community, is used here since the behavior can be 59
4 thought of as a re#ex. This is a reasonable safety strategy 60
5 since the robot may be moving around in a highly un- 61
6 structured world, where the occurrences of unpredicted, 62
7 or unmodeled obstacles are very likely. 63
8 We assume that we can control the robot's transla- 64
9 tional and rotational velocities. Naturally, for platforms 65
10 that do not give direct control over the velocities, one 66
11 needs to design an actuator control so that these velocity 67
12 controls are realized. The implementation could be just 68
13 a static mapping, as in the car case, or a dynamic regula- 69
14 tor. In those cases, the velocity controllers can be viewed 70
15 as higher-level controllers. 71
16 An example of systems that give direct velocity control Fig. 1. The two possible goal-attraction and obstacle-avoidance auto- 72
mata. (a) Concurrent behaviors, (b) Hard switches.
17 is the Nomadic 200 mobile platform, which is the experi- 73
18 mental platform used in this work. For such systems, 74
19 a unicycle model (see for example Ackermann, 1993) is 75
20 quite standard: 76
21 for example Lygeros, Tomlin, & Sastry, 1999) for describ- 77
22 x
"v cos
, ing the situation. This depends on whether the two be- 78
23 haviors are active simultaneously or not, as seen in Fig. 1, 79
y
"v sin
, (1)
24 where z"(x, y,
) and where we let f (z) and f (z) 80
- %
25
Q ", denote the full state, obstacle-avoidance and goal-attrac- 81
26 tion behaviors, respectively. Initially we let v be constant 82
27 where (x, y)31 is the position of the robot,
is its and we postpone the discussion about how to design v to 83
28 heading, and v and  are the controlled translational and Section 4.2. 84
29 rotational velocities, respectively. Furthermore, if the so- If one chooses to work with concurrently active behav- 85
30 nars on the robot detect a point-obstacle at (x , y ) iors, di!erent controllers a!ect the system simulta- 86



31 that is closer to the robot than the prede"ned safety neously, resulting in a smooth overall performance, as 87
32 distance d , a standard, reactive control response is shown in Kos\ eckaH (1996). However, in this case the 88
-
33 analysis of the system is becoming signi"cantly more 89
34 "C = (d) (
I !
), (2) di$cult as new behaviors are added. 90
- -
35 The other possible solution to the coordination prob- 91
36 where d"((x!x )#(y!y ), = (d)"1/d if lem, corresponding to hard switches between the di!er- 92


-
37 d(d and 0 otherwise, and
I "#a tan 2(y !y, ent behaviors, has the major disadvantage that it both 93
- 

38 x !x). Here, C is a constant weight, and the choice a!ects the performance in a negative way, not allowing 94

-
39 of C should re#ect the hardware constraint on the for the smooth performance that concurrently active 95
-
40 maximal angular velocity. The choice of d should behaviors produce, and that it increases the risk of intro- 96
-
41 re#ect the sensitivity and accuracy of the sensors. ducing chattering into the system. Therefore our idea is 97
42 to impose hard switches on the behavior based system in 98
43 such a way that we can model each behavior as a node in 99
44 3. A hybrid control approach an automaton. The reason for using hard switches is due 100
45 to the fact that we want each individual behavior to 101
46 When adding a goal attraction behavior, de"ned in correspond to a node in the automaton, due to scalability 102
47 the same way as the obstacle-avoidance behavior except issues. This construction also enables us to theoretically 103
48 that we now have an attractive instead of a repulsive analyze the performance of each behavior individually. 104
49 "eld, we get two di!erent possible hybrid automata (see At the same time, we want to avoid the negative, chatter- 105
50 ing e!ects that such an approach could potentially give 106
51 rise to. This will be done by adding nodes to the automa- 107
52  Since a real, extended obstacle cannot be considered to be a point, ton as a way of regularizing it. In what follows, we will 108
53 in the actual implementation of the avoidance behavior, the desired show that even though we introduce hard switches, the 109
heading needs to be calculated as the orientation of the sum of the
54 weighted vectors that each individual sonar reading contributes with.
performance is not a!ected much when using a regular- 110
55 For a Nomadic 200 this corresponds to taking the sum over 16 ized automaton instead of concurrently active behaviors. 111
56 elements since that platform is equipped with 16 ultrasonic sensors. In other words, what we want to do is to eliminate the
 M. Egerstedt, X. Hu / Automatica 38 (2002) 125}130 127

1 57
2 58
3 59
4 60
5 61
6 62
7 Fig. 2. Goal attraction together with obstacle-avoidance results in 63
8 a Filippov type Zeno automaton. The grey region around the obstacle 64
9 is the region where obstacle-avoidance is active. The arrows correspond 65
10 to the di!erent vector "elds that are acting on the robot. 66
11 67
12 68
13 possibility of the so-called Zeno phenomena in the sys- 69
14 tem. What this corresponds to is a hybrid system that 70
15 exhibits an in"nite number of discrete transitions in "nite 71
16 time (Johansson, Egerstedt, Lygeros, & Sastry, 1999). 72
17 For the type of automata that we encounter here, the 73
18 in"nite number of discrete transitions, made in "nite 74
19 time, is caused by the fact that the underlying system that Fig. 3. Simulation of (a) concurrently active behaviors, (b) hard 75
switches, and (c) a regularized automaton on the Nomadic simulator,
20 the automaton tries to model is a switched system that the Nserver.
76
21 exhibits sliding in the sense of Filippov (1988). They thus 77
22 form a special class of Zeno hybrid automata since they, 78
23 theoretically, make an in"nite number of transitions in 79
24 zero time. Such sliding modes arise when the systems where 3[0,1] is chosen so that f is orthogonal to 80
1
25 have continuous #ows that point toward the switching f . Adding this type of information about the 81
-
26 surface, resulting in a new, induced #ow on that surface. di!erent behaviors makes it possible to generate the extra 82
27 In these cases, the automaton can be regularized by the node in the automaton automatically. It furthermore 83
28 introduction of a new node with the continuous #ow suggests that our method would scale when more than 84
29 given by the Filippov solution (Malmborg, 1998). two behaviors a!ect the motion of the robot, as 85
30 If we now assume that C in (2) is large enough so long as an automatic procedure for designing the sliding 86
-
31 that the heading of the robot can be considered to be solutions could be identi"ed for the new behaviors as 87
32 more or less instantaneously driven to its desired con"g- well. 88
33 uration, the hybrid automaton in Fig. 1(b) can admit The assumption about instantaneous heading control 89
34 Zeno executions. This obvious fact is best illustrated by is obviously a simpli"cation but it still gives a model that 90
35 Fig. 2, where the extra node that needs to be added in is rich enough to capture the, from our point of view, 91
36 order to regularize the automaton can easily be identi"ed relevant phenomena. In fact, in real life we have a possi- 92
37 as well. This extra node is just a node containing the bility of chattering that here reveals itself as a Zeno 93
38 sliding dynamics that is de"ned on the boundary between execution. 94
39 the two behaviors. The regularized point-to-point automaton was imple- 95
40 When an obstacle is closer to the robot than d , the mented and tested on the Nomadic 200 mobile robot. 96
-
41 obstacle-avoidance behavior becomes active. Since the In Fig. 3, the results from running the system on 97
42 repulsive potential "eld from that behavior is orthogonal the Nserver, the Nomadic simulation package, 98
43 to the surface on which the behavior becomes active, the can be seen. In Fig. 3(a), concurrently active behav- 99
44 sliding solution is just iors are displayed, resulting in a smooth movement 100
45 around the obstacles, while the chattering solution 101
46 f "f #(1!) f , (3) in Fig. 3(b) corresponds to hard switches. The reason 102
1 - %
47 why we do not have sliding in this case is due to the 103
48 part of the dynamics of the robot that was ignored 104
49  The name Zeno refers to the philosopher Zeno of Elea (500}400
in the analysis. It is still clear that from a perfor- 105
50 BC), whose major work consisted of a number of famous paradoxes. mance perspective, Fig. 3(b) is an unsuccessful 106
51 They were designed to explain his view that the ideas of motion and design. In Fig. 3(c) the result from using a regularized 107
52 evolving time lead to contradictions. An example is Zeno's Second hybrid automaton can be seen, and even though we only 108
53 Paradox of Motion, in which Achilles is racing against a tortoise. 109
 The other class of Zeno automata has a slightly more complex
54 dynamics. Here the automaton changes nodes faster and faster, with the
110
55 jump times converging to a "nite, so-called Zeno time (Johansson et al.,  This typically depends on whether we have access to a geometric 111
56 1999). description of the switching surface or not.
 128 M. Egerstedt, X. Hu / Automatica 38 (2002) 125}130

1 have one behavior active at a time, the performance is 57

2 satisfactory. 58
3 59
4 60
5 4. Path planning and tracking 61
6 62
7 Given the reactive obstacle-avoidance behavior from 63
8 the previous section, the main question that we want to 64
9 address here is: How do we construct an appropriate 65
10 approach-target behavior? Obviously, we can do better 66
11 than to just use an attractive potential "eld, and it will 67
12 turn out that our regularized automata approach allows 68
13 us to answer this question. What we want to do is to 69
14 produce a robot behavior that satis"es the safety speci- Fig. 4. A heuristic, near-optimal high level control design is shown. 70
15 "cations at the same time as the solution is close to 71
16 optimal with respect to a given performance cost func- 72
17 tion. A "rst formulation, inspired by Tomlin, Papas, 73
18 Kos\ eckaH , Lygeros, and Sastry (1998), of what we want to The main idea is that instead of focusing on the hard 74
19 accomplish could be the following: If we let our admiss- optimal control problem, we should concentrate on pro- 75
20 ible controls be u3;, and assume that we only encounter ducing suboptimal (but close to optimal) geometric paths 76
21 one point obstacle (x , y ), we can de"ne a safety cost that lead around the obstacles. This way we do not have 77



22 function to deal with the actual kinematics of the robot in the 78
23 optimization formulation. Instead, we design our track- 79
24 J (u)"min (x(t)!x )#(y(t)!y ), (4) ing controller in such a way that the robot can even track 80
 


25 RV reasonably well those paths which do not always meet 81
26 where the dependence on the control, u, is given impli- the kinematical constraints. This means that we cannot 82
27 citly by the controlled system dynamics from the pre- be sure that we actually "nd the optimal controller, but 83
28 vious section. The set of controls, ; (C), that make the rather that we "nd one that is reasonably close to the 84

29 robot move at least a distance C away from the obstacle optimal one as long as we have a good enough trajectory 85
30 can thus be de"ned as ; (C)"u3;: J (u)*C. tracker. 86
 
31 The next step is to de"ne another cost function that The desired overall behavior that these heuristics give 87
32 penalizes high curvature of the chosen path. This is rise to (under the assumption of perfect tracking), to- 88
33 a reasonable performance criterion since it penalizes gether with the corresponding automaton, is depicted in 89
34 paths that make the robot move in sudden, abrupt ways. Fig. 4, where an optimal path is planned and followed by 90
35 Furthermore, this smoothness objective gives a trajectory an approach-target behavior until an obstacle is detec- 91
36 that a robot has good chances of following when it is ted. Then an approach-obstacle behavior follows another 92
37 governed by physical limits on what signals the actuators path to the region where the regularized sliding behavior 93
38 can actually track. becomes active. When the target can be reached by an 94
39 The idea now is to choose the control candidates for optimal path, not intersecting the safety region around 95
40 minimizing this new performance cost function from the the obstacle (called detect-target in the "gure), ap- 96
41 set of safe controls, ; (C ), where C is our preferred proach-target becomes active again. 97
  
42 safety margin. Unfortunately, it turns out that this is 98
43 a very hard problem to solve numerically (not to mention 99
44 analytically), as pointed out in Tomlin et al. (1998). It is 4.1. Path planning 100
45 thus not suitable in situations where on-line computa- 101
46 tions are necessary. However, the underlying approach One "rst observation is that for a path produced 102
47 could suggest a way for producing a solution to the by a scalar function y "f (x ), the curvature is given 103
 
48 obstacle negotiation problem that is both safe, com- by 104
49 putationally feasible, and makes the system behave in 105
50 a satisfactory way with respect to keeping the curvature f (x ) 106
51 of the produced path small. (x )"  , (5) 107
 (1#f (x ))
52  108
53 109
54  It should be mentioned that both J and ; depend on the robot's
where the subscript d stands for the desired robot posi- 110
55   tion, and f ( ) ) and f ( ) ) denote "rst and second deriva- 111
initial position, but for the sake of notational simplicity we leave that
56 out from the de"nitions. tives, respectively.
 M. Egerstedt, X. Hu / Automatica 38 (2002) 125}130 129

1 Thus, if we minimize f (x ) instead of (x ) we make perfectly, we would have 57

 
2 (x ) small automatically, which is a desired feature, as 58
 p (s) q (s)
3 seen in the previous paragraph. s
" x
# y
(8) 59
4 Since we, by following this proposed route, minimize p (s)#q (s) p (s)#q (s) 60
5 the ¸-norm of the second derivative, the resulting curve since this corresponds to x
"x
and y
"y
. On the 61
6 will be a cubic spline. This is a fortunate fact since it   62
other hand, (8) does not contain any position error
7 means that we will not be forced to relay on extensive feedback, which is important for the robustness. There- 63
8 world information or to do any heavy computations fore we propose our dynamics for the reference point as 64
9 on-line which tends to be the case when more sophisti- follows: 65
10 cated planning algorithms are used (Krogh & Thorpe, 66
11 1986; Latombe, 1991; Stenz, 1994). It is thus an almost ce\?Mv 67
s
"  , (9)
12 trivial task to generate the splines that connect the robot (p (s)#q (s) 68
13 and the target in the approach-target behavior, and the 69
where v is the desired speed at which one wants the
14 robot and the obstacle in the approach-obstacle behav-  70
15 ior, as seen in Fig. 4. vehicle to track the path, and  and c are appropriate, 71
16 positive numbers. 72
17 4.2. Tracking We now let our control algorithm be as follows: 73
18 v"
cos(e ) 74
19 We now have an on-line method for producing low ( 75
"ke #
Q , k'0, (10)
20 curvature paths around detected obstacles, and hence (  76
21 our next task is to "nd a good tracking algorithm so that where both and k are positive, e "
!
, and 77
22 (  78
the robot follows the proposed path robustly.
"arctan 2(y !y, x !x).
23 We let the general reference path, parameterized by s,    79
24 be given by Remark. We should point out that e is not de"ned at 80
25 ( 81

"0 since
is not de"ned. In implementation, one can
26 x "p(s),
  82
0)s)s , (6) replace
by
27 y "q(s), D  83


28
where the idea is to let the motion of the reference point  84
29
be governed by a di!erential equation containing error if
', 85

I "
30
feedback. It can be viewed as a combination of the 
(!2
#3
)# (!2(!
)#3(!
)) 86
31  P 87
conventional trajectory tracking, where the reference tra- 
32 if
), 88
jectory is parameterized in time, and a dynamic path
33 89
following approach (Sarkar, Yun, & Kumar, 1993), where (11)
34 90
the criterion is to stay close to the geometric path, but not
35 where  is a small positive number. It is easy to 91
necessarily close to an a priori speci"ed point at a given
36 see that
I is well de"ned at
"0 since 92
time. This approach makes our algorithm robust to 
37 lim
(!2
#3
)"0. 93
measurement errors and external disturbances since, if M 
38 94
both the tracking errors and disturbances are within
39 An intuitive interpretation for this control algorithm is 95
certain bounds, the reference point moves along the refer-
40 that the robot is steered toward the reference point on the 96
ence trajectory while the robot follows it within a pres-
41 desired path with a speed proportional to the tracking 97
peci"ed look-ahead distance. Otherwise, the reference
42 error and whenever it reaches the steady tracking error 98
point should slow down and `waita for the robot, as
43 (the desired looking-ahead distance) it will track the path 99
suggested in Egerstedt, Hu, and Stotsky (1998)
44 with an almost constant speed. It is easy to see that even 100
Our control objectives are
45 if a path is planned beyond the dynamic or kinematic 101
46 lim sup
(t)) constraints of the system, this control algorithm can still 102
47 M be used to track the path the best it can. 103
R
48 In Egerstedt et al. (1998), it was shown that for the 104
lim sup 
(t)!
(t)) , (7)
49  ( platform model (1), governed by the control (10), the 105
50 R steady state tracking error,
, can be made as small as 106
51 where  and  are positive numbers that can be made one wants while
tends to
exponentially. Further- 107
M ( 
52 arbitrarily small,
(t)"((x !x)#(y !y), where more, in steady state we have that v+v . Thus we, by 108
  
53 (x, y) is the actual position of the robot, and
and
are using the control law (10), meet the control objectives 109

54 actual and desired robot orientations. de"ned in (7). 110
55 From (6) we directly get that x
"p (s)s
, y
"q (s)s
, We thus have a way of both producing and tracking 111
 
56 which implies that if the robot would track the path paths, and we now combine these two together into the
 130 M. Egerstedt, X. Hu / Automatica 38 (2002) 125}130

1 Filippov, A. F. (1988). Diwerential equations with discontinuous righthand 57

2 sides. Dordrecht: Kluwer Academic Publishers. 58
3 Johansson, K., Egerstedt, M., Lygeros, J., & Sastry, S. (1999). Regulariz- 59
ation of Zeno hybrid automata. Systems and Control Letters, 38,
4 141}150.
60
5 Kortenkamp, D., Bonasso, R.P., & Murphy, R. (Eds). (1998). 61
6 Artixcial intelligence and mobile robots. Cambridge, MA The MIT 62
7 Press. 63
8 Kos\ eckaH , J. (1996). A framework for modeling and verifying visually 64
guided agents: Design, analysis and experiments. Dissertation, Grasp
9 Fig. 5. Implementation of the control approach from Fig. 4. The robot Lab, March.
65
10 tries to reach a goal point straight ahead of it but is forced to negotiate Krogh, B., & Thorpe, C. (1986). Integrated path planning and dynamic 66
11 unmodeled obstacles and walls. steering control for autonomous vehicles. Proceedings of the 1986 67
12 IEEE International Conference on Robotics and Automation, San 68
13 Francisco, CA (pp. 1664}1669). 69
Latombe, J. C. (1991). Robot motion planning. Dordrecht: Kluwer
14 path following behavior that moves the robot safely Academic Publishers.
70
15 around the obstacles at the same time as its executed Lygeros, J., Tomlin, C., & Sastry, S. (1999). Controllers for reachability 71
16 trajectories are not too far from optimal with respect to speci"cations for hybrid systems. Automatica, 35(3), 349}370. 72
17 curvature. As seen in Fig. 5, where real experimental data Malmborg, J. (1998). Analysis and design of hybrid control systems. Ph.D. 73
18 are displayed, the method seems to work well. thesis, Department of Automatic Control, Lund Institute of 74
Technology, Lund, Sweden, May.
19 Sarkar, N., Yun, X., & Kumar, V. (1993). Dynamic path following:
75
20 A new control algorithm for mobile robots. Proceedings of the 32nd 76
21 5. Conclusions Conference on Decision and Control, San Antonio, Texas, 77
22 December. 78
23 Stenz, A. (1994). Optimal and e$cient path planning for partially 79
In this paper, it is shown that a behavior based control known environments, Proceedings of the 1994 IEEE International
24 system can be modeled as a hybrid automaton, where Conference on Robotics and Automation, San Diego, California.
80
25 each node corresponds to a distinct robot behavior. In Tomlin, C., Papas, G., Kos\ eckaH , J., Lygeros, J., & Sastry, S.S. (1998). 81
26 order to achieve this, we have to impose hard switches on Advanced air tra$c automation: a case study in distributed decen- 82
27 tralized control. Control problems in robotics, Lecture Notes in Con- 83
the transitions between the di!erent behaviors, resulting trol and Information Sciences, Vol. 230. London: Springer.
28 in a potentially chattering overall behavior. We further- 84
29 more show how regularization techniques can be used 85
30 to solve this problem by adding extra nodes to the 86
31 Magnus B. Egerstedt was born in Stock- 87
automaton. Those extra nodes correspond to the sliding holm, Sweden, in 1971. He received the
32 dynamics on the boundary between the di!erent M.S. degree in Engineering Physics and 88
33 behaviors. The performance aspect of this approach the Ph.D. degree in Applied Mathematics 89
34 from the Royal Institute of Technology, 90
is veri"ed experimentally on a Nomadic 200 mobile Stockholm, in 1996 and 2000, respectively.
35 platform. He also received a B.A. degree in Philos- 91
36 We also propose a method for designing reach-target ophy from Stockholm University in 1996. 92
37 He is an Assistant Professor in Electrical 93
behaviors in such a way that questions concerning safety and Computer Engineering at the Georgia
38 and curvature minimization can be addressed explicitly. Institute of Technology. 94
39 Our proposed, suboptimal method is based on a combi- He spent 2000}2001 as a Postdoctoral 95
40 Fellow at the Division of Engineering and Applied Science at Harvard 96
nation of path planning and trajectory tracking tech- University and during 1998 he was a Visiting Scholar at the Robotics
41 niques, placing it in the deliberative part of the behavior Laboratory at the University of California, Berkeley. Dr. Egerstedt's 97
42 based control architecture spectrum. Furthermore, we research interests include optimal control as well as modeling and 98
43 analysis of hybrid systems, with emphasis on motion planning and 99
show that this approach works well in practice on our control of mobile robots.
44 experimental platform. 100
45 101
46 Xiaoming Hu was born in Chengdu, 102
47 China, in 1961. He received the B.S. degree 103
References from University of Science and Techno-
48 104
logy of China in 1983. He received the M.S.
49 Ackermann, J. (1993). Robust control. London: Springer. and Ph.D. degrees from Arizona State 105
50 Arkin, R. C. (1998). Behavior-based robotics. Cambridge, MA: The MIT University in 1986 and 1989 respectively. 106
51 Press. From 1989 to 1990 he was a Gustafsson 107
Postdoctoral Fellow at the Royal Institute
52 Brooks, R. (1986). A robust layered control system for a mobile robot. 108
of Technology, Stockholm, where he is
53 IEEE Journal of Robotics and Automation, 2(1), 14}23. currently an associate professor. His main 109
Egerstedt, M., Hu, X., & Stotsky, A. (1998). Control of a car-like robot research interests are in nonlinear feed-
54 using a virtual vehicle approach. Proceedings of the 37th IEEE
110
back stabilization, nonlinear observer de-
55 conference on decision and control, Tampa, FL, USA, December sign, sensing and active perception, motion planning and control of 111
56 (pp. 1502}1507). mobile robots, and mobile manipulation.