[springer tracts in advanced robotics] distributed autonomous robotic systems volume 83 ||...

Cooperative Transportation by Swarm RobotsUsing Pheromone Communication

Ryusuke Fujisawa, Hikaru Imamura, and Fumitoshi Matsuno

Abstract. Ants communicate with each other using pheromones, and their societyis highly sophisticated. When foraging, they transport cooperatively with interplayof forces. The swarm is robust against changes in internal state, and shows flexi-bility in dealing with external problems. In this brief paper, we focus on the robotswarm that achieves cooperative transportation making use of ethanol as a substan-tial artificial pheromone. We also propose a swarm system with a newly developedalgorithm that enables cooperative transportation of real robots. They will transportfood to the nest analogous to the behaviour of a swarm of ants. Emphasis will beplaced on the systematic task solution process. We present a number of experimentsdemonstrating the robustness and flexibility of the system and also confirming theeffectiveness of the algorithm.

1 Introduction

1.1 Basic Characteristics of a Swarm

Generally, a swarm is a distributed autonomous system. It acts only according tolocal information in the given environment without any global information. An in-dividual acts autonomously in the swarm, according to the circumstances [1]. Globalbehaviour emerges by interactions among individuals. Thus, these interactions are

Ryusuke FujisawaHachinohe Institute of Technology, Hachinohe, Japane-mail: [email protected]

Hikaru ImamuraDENSO, Japan

Fumitoshi MatsunoKyoto University, Kyoto, Japane-mail: [email protected]

A. Martinoli et al. (Eds.): Distributed Autonomous Robotic Systems, STAR 83, pp. 559–570.springerlink.com c© Springer-Verlag Berlin Heidelberg 2013

[email protected]

[email protected]

560 R. Fujisawa, H. Imamura, and F. Matsuno

fundamental term for formation of the swarm. Living organisms that form swarmscommunicate with each other to interact frequently. Therefore, swarms have robust-ness -a property for adapting to changes in the internal state- and also flexibility -aproperty for adapting to changes in the external state (e.g., the environment)[2].

1.2 Pheromone Communication in Ants

The social insects, such as ants and termites, are known to communicate with eachother and form swarms using pheromones [3]. Ants form especially complex soci-eties [3, 4, 5].

A pheromone is any chemical or set of chemicals produced by a living organ-ism that transmits a message to other members of the same species [6]. In this pa-per, we focus on foraging behaviour of ants using a pheromone. When an ant findsand brings food back to the nest, it secretes a pheromone that forms a trail. Theother ants trace the pheromone trail and reach the food. An ant stops to lay downthe pheromone trail when it cannot find the food. The pheromone trail accordinglyvolatilises and/or diffuses into the environment, and thus information meaninglessto the ants disappears. This is a simple but advanced communication method.

Itou et al. reported that the merits of this method are: 1) local and decentralised in-formation management, and 2) self-propagation effect for information exchange [7].Thus, pheromone communication is a suitable method for a distributed autonomousrobot system. A number of individuals can converge at the food, communicatingwith each other. As each individual shares the purpose of action, i.e., collectingfood, there is an interplay of forces that drive cooperative transportation. Each in-dividual acts in accordance with a very simple corrective model, but the swarmshows advanced behaviours. Thus, “pheromone communication in the ant colony”and “emergence of cooperative transportation by interplay of forces” are useful toapply to robotics. Once we provide swarm robots with functions of task solution,the robots spontaneously find a method by interaction with each other, which is ofremarkable significance.

1.3 Related Studies and Issues

Several studies, such as those of Sugawara et al. [8] and Garnier et al. [9], havedemonstrated swarms of robots achieving foraging behaviour of ants with a vir-tual pheromone. In these studies, however, the swarm is inevitably non-autonomousin that it requires an external measurement system composed of a projector and acamera. Using substantial pheromones to control the robots is required to make theswarm autonomous.

Shimoyama et al. [10] achieved pheromone tracking behaviour using real in-sect antenna and substantial pheromone, but biomaterials cannot be handled easilyby swarm robots. In addition, Shimoyama et al. did not pay particular attention tothe swarm behaviour. A number of problems remain to be resolved in pheromone

Cooperative Transportation by Swarm Robots Using Pheromone Communication 561

communication robotics. Diffusion, an important factor for pheromone communica-tion, is just one of these problems. To adjust the duration of the pheromone signal, itis necessary to change the concentration of a pheromone and/or mix with some othersubstance(s). In addition, only a few advantageous chemical sensors are availableat present. Purnamadjaja et al. [11] studied swarm robots that communicate usingtwo chemical substances, regulating a gas sensor in a sophisticated manner. How-ever, as only one robot secretes the pheromone, this system achieves only one-sidedcommunication.

There have been some studies on cooperative transportation by swarm robots.Dorigo et al. [12] developed a swarm robot system, “Swarm-bots”, in which eachrobot has a grasping mechanism connecting the individual robots, which enables therobots to run through a gap or a trench. They accomplish cooperative transportationwith the interplay of forces [13]. The robot itself and/or packages to be transportedemit light, and the robots recognise what they should transport. Kube et al. alsoaccomplish cooperative transportation of swarm robots [14]. The robots recognisethe target to be transported with the installed light receiving element. These studiesfocused on the interplay of forces, but did not pay a great deal of attention to theindirect communication to advance emergence.

In our previous study, we achieved pheromone communication of swarm robotsfor recruitment behaviour [1]. In the present study, we propose cooperative trans-portation using pheromone communication as seen in ants. Experiments with thenewly developed swarm robots indicated the effectiveness of pheromone communi-cation in cooperative transportation, and suggested the robustness and flexibility ofthe swarm.

2 Swarm Behavior Algorithm

2.1 Swarm Behavior as Deterministic Finite Automaton

We have also been investigating a transportation algorithm. The result of previousstudies [1, 15, 16, 17] will be applied to the new algorithm logic presented below. Inprevious studies, we focused on pheromone communication. Here, we concentrateon cooperative transportation with pheromone communication as seen in ants. Thealgorithm is described by the deterministic finite automaton shown in Fig. 1. Therobots act in a completely autonomous manner by this algorithm. This algorithmpremises that swarm robots search for food in a given field, and transport the foodto the nest.

To design this algorithm, we defined 6 internal states, Si(i = 1, · · · ,6); 10 percep-tual cues (stimuli), Pi(i = 1, · · · ,10); and 6 effector cues (actions), Ei(i = 1, · · · ,6).We also assumed that there are many robots in the field, and that all agents can de-tect the direction of the nest as in the case of ants. As shown in Fig. 1, the agentwhose state is Si selects the action Ei. If the agent in state Si detects the perceptual


S1

E1

S2

E2

S3

E3

Collisionprocessing [3]



P4P3

P9

P10

P9

P10

P9

P10

Random walk

Attraction state

Laying down pheromone

Tracing state

Tracing pheromone trail

Timeout occurs

Arriving to the nest

Presence of pheromone

Contact with other object



Complete processing

S4

E4

S5

E5


Initial state

Going around the food

Transport state

Pushing the food

Preparing transport state

Arriving to the nest

Finding food

P1

Finding food

Losing food

P6P7 Complete preparation

Complete processing

Complete processing

Un-transportable

Changing direction

S : Internal state [i = 1, ,6] (State)

Ei : Effector cue [i = 1, ,6] (Behavior when state i)Collision

processing [i ]: Processing for collision withanother objects when state i ( i = 1, ,4,6): Perceptual cue [j = 1, ,10] and direction of state transitionP

P5 Losing food

P9

P10



Complete processing

S6

E6

Preparing attraction state

Going around the foodP6

Complete preparation

P8P2

P9

P10


Complete processing

P2

P1

P5

Fig. 1 Algorithm for cooperative transportation using pheromone communication

cue Pj, the state of the agent is transited to Sk. The details of the internal states Si ofthe robot, perceptual cues Pi and effector cues Ei are as follows.

Si: S1, Search: the agent does not have any information on the food; S2, Attrac-tion: the agent has the location information on the food; S3, Tracing: the agent hasonly the direction information on the location of the food; S4, Pre-transportation:the agent has the location relationship of the nest and the food. As each robot canonly push (not pull), it needs to run around the food so that the latter is on the linebetween the former and the nest; S5, Transportation: the robot pushes the food; S6,Pre-attraction: when the food will not move, the robot runs around it to find the wayand returns to the nest.

Pi: P1, Contact with food; P2, Nest arrival; P3, Presence of pheromone; P4, Time-out occurrence; P5, Losing pheromone trail; P6, Completion of running around thefood; P7, Necessity of direction adjustment during transportation; P8, Impossibilityof transportation; P9, Contact with other object (a wall or other robots); P10, Com-pletion of collision processing.

Ei: E1, Random walk; E2, Pheromone secretion; E3, Following the pheromonepath; E4, Running around the food (to get behind it and on the line between the foodand the nest); E5, Pushing the food; E6, Running around the food (to go to the nest).

2.2 Collision Processing

The robot perceives its external environment by contact with object(s), and acts inaccordance with the collision algorithm shown in Table 1. The improvement of thepreviously developed collision algorithm [1] enabled us to avoid traffic jams on thepheromone trail. In addition, as the robots always make contact with each other dur-ing cooperative transportation, collision processing is not executed at the state S5.


Table 1 Behaviour selection of algorithm for cooperative transportation after collision

Collision Processing Contact position Robot’s behaviour after making contact with other object

Collision Processing [1]front rotation on-site after disengaging from contact point by reversingback rotation on-site after disengaging from contact point by proceeding

Collision Processing [2]front stop on-siteback disengaging from contact point by proceeding

Collision Processing [3] entire circumference disengaging from contact point by reversing

Collision Processing [4]front disengaging from contact point by reversingback disengaging from contact point by proceeding

Collision Processing [6]front disengaging from contact point by reversingback disengaging from contact point by proceeding

Nest sensor (IR phototransistor)

Micro pump for laying down pheromoneEthanol tank

Alcohol sensor

Internal state transmitter

DC Motor & magnetic encoder

Touch sensor & LED – Cds unit

Nest sensor (IR phototransistor)

Micro pump for laying down pheromoneEthanol tank

Alcohol sensor

Internal state transmitter

DC Motor & magnetic encoder

Touch sensor & LED – Cds unit

Fig. 2 Construction of the robot developed here (ARGOS01)

3 Construction of Robot

Figure 2 shows our newly developed swarm robot, ARGOS01. The robot has twoactive wheels and four castors. Its active wheels can be controlled independentlyso that the robot can move on a flat plane. The specifications of ARGOS01 areas follows: body diameter, 150 [mm]; height, 195 [mm]; weight, 1.26 [kg]; andmaximum speed, 0.1 [m/s]. It has a Ni-MH battery (7.2 [V], 3900 [mAh]) at itscentre. Four microcontrollers (Cypress Semiconductor) are installed in the robot. Itsmaster controller is connected to 2 slave controllers by I2C, and these three processthe data from the sensors and control the motors. The other microcontroller sendsthe internal state to an external computer, which is used only to observe the internalstate and swarm behaviour of the robots.

To implement the algorithm described in 2.1, a robot needs sensors to detect10 perceptual cues (Pi (i = 1, · · · ,10)), and actuators or mechanisms to carry out 6


effector cues (Ei (i = 1, · · · ,6)). The following are the sensors and actuators installedon the robot.

• Nest sensor: detects the direction of the nest with infrared lamps.• Touch sensor: detects contact with other objects.• Photoreceptor unit: detects food (the food emits light by blue LED).• Rotary encoder: detects success/failure of transportation.• Alcohol sensor: detects pheromone.• Pheromone secretion mechanism: lays down a pheromone trail.

The robot detects a collision with push switches at the side of its body. The foodhas blue LEDs at all around the body. With its photoreceptor unit, the robot looksfor an object with LED light, and identifies it as food at contact. Based on the re-sults of the previous experiments, we set the perceptible distance of the robot asaround 300 [mm]. Rotary encoders at motors are used to detect the transport state.In this study, instead of a biological pheromone, we used ethanol, which is a volatilesubstance similar to the trail pheromone. The robot has alcohol sensors to detectthe pheromone trail on the experimental field. A micropump on the robot secretesethanol from the installed 50 [ml] tank.

The robot is composed of four layers supported by spacers. The first (bottom)layer has DC motors with a rotary encoder, alcohol sensors and an ethanol vent.The system substrate, the sensor substrate for the photoreceptor units and touchsensors, and the power and wireless communication substrate are installed at thesecond layer. The third layer has a micropump and a tank to secrete ethanol. Thefourth (top) layer has the nest sensor to determine the direction of the nest.

4 Experiments and Results

Applying the designed swarm behaviour algorithm to the robot system, we have ver-ified the effectiveness of pheromone communication in cooperative transportation.Figure 3 shows the experimental field: a 3,600 [mm]×1,800 [mm] 2D flat planesurrounded by walls. The diameter of the food is 300 [mm] and that of the nest is900 [mm]. Considering the perceptible distance of the robot, we set as the perceiv-able area of the robot a circle of radius 450 [mm] with the food at its centre. If therobot goes into the perceivable area, it can detect the direction of the food and makecontact. As this study focused on cooperative transportation by the swarm robots,the weight of the food should be heavy enough so that one robot cannot move itby itself; we set the weight as 3.58 [kg]. This weight requires the cooperation of atleast three robots. The transportation distance is 2,000 [mm]. We define a task solu-tion when the robots transport the food to the nest. To trace the pheromone (100%ethanol) trail, we set a threshold on a value detected by the alcohol sensor. We setthe signal duration time to trace the pheromone trail as 3 minutes. As a result ofthese settings, the robot continues to trace the pheromone trail for 3 minutes. Therobot can lays down 5 trails in an experiment. When tank is empty, we change therobot with reserve robots. If the robot depletes the pheromone (ethanol), we change


the robot to continue the experiment. We also define that the robot recognises the P8

(impossibility to transport) 20 [s] after starting pushing the food.Figure 4 shows the results of an experiment with 10 robots. A, B and C are nor-

mal camera images; and A’, B’ and C’ are thermographic images, which allow usto distinguish the pheromone trail with its lower temperature caused by heat evap-oration of ethanol. The robots had laid down the pheromone trail at 10 [min]. At20 [min], they had finished transporting the food to the nest in a concerted manner.The results of the experiment clearly showed both that the robots achieved coop-erative transportation using the pheromone trail and that our new swarm behaviouralgorithm works effectively. In the next chapter, setting the task solution time as theevaluation index, we will consider the robustness and flexibility of the swarm.

4.1 Effect of Pheromone Communication

We performed an experiment to determine the effectiveness of the pheromone com-munication on cooperative transportation. Figure 5 shows the relationship betweentask solution time [with or without pheromone communication] and the numberof individuals; the vertical axis represents the average task solution time of 10 tri-als, and the horizontal axis represents the number of robots used. Error bars showthe maximum and minimum in the experiments. With four or seven robots, thetask solution time with pheromone communication was shorter than that withoutpheromone. The pheromone communication was effective when the density of therobots was low in the field. With a high density (10 robots) in the field, pheromonecommunication did not strongly influence the task solution time. This indicated thatpheromone communication is not always necessary for cooperative transportation

450 mm900 mm

300 mm

2000 mm

3600 mm

900 mm

1800 mm

Food

Food perceptible area

Nest

Transport distance

150 mm

Robot

Fig. 3 Plane view of experimental field


0 min

0 min

A

A’

Food

Robot

10 min

10 min

B

B’

PheromonetrailNest

20 min

20 min

C

C’

Fig. 4 Snapshots of experimental result by 10 robots using pheromone communication(A,B,C : Shot by digital cameraCA’,B’,C’ : Shot by thermography)

0

500

1000

1500

2000

2500

3000

4 7 10

Robot number

Task c

om

ple

te t

ime [

sec]

Not using pheromone communication

Using pheromone communication

0

500

1000

1500

2000

2500

3000

4 7 10

Robot number

Task c

om

ple

te t

ime [

sec]

Not using pheromone communication

Using pheromone communication

Fig. 5 Experimental result comparing cases with and without pheromone communication

when a sufficient number of robots are in the field as many robots find the food andbegin cooperative transportation before laying down the pheromone trail.

Table 2 shows the average time of each event of 10 trials. When only a few robots(4 robots) are in the field; we found a marked difference between pheromone andnon-pheromone communications. Without the pheromone communication, as therobots depend completely on the random walk to congregate at the food, they takea great deal of time to start pushing the food. With pheromone communication, itattracts the robots and they can begin to transport the food quickly. However, therewere no clear effects of pheromone communication with ten robots in the field forthe same reason as described above for the case of the task solution.


Table 2 Comparing average times of transportation starting in cases with and withoutpheromone communication

number of robotspheromone 4 robots 7 robots 10 robotsWith 625 [s] 335 [s] 361 [s]Without 1057 [s] 612 [s] 355 [s]

Fig. 6 Experimental result for robustness (percentage of disable robots : 2/7)

4.2 Robustness of the Swarm

To demonstrate the robustness of the swarm robots, we performed an experimentsimilar to that described above in 4.1 with seven robots using pheromone commu-nication. During cooperative transport, we stopped two robots (S5) in the swarm todetermine whether the robots keep functioning as a swarm even when the swarmloses its soundness.

Figure 6 shows the typical experimental result. The vertical axis represents thepercentage of internal states of the robots in the swarm, and the horizontal axis rep-resents the experiment time. Light grey indicates that the robot is in state S1 (search),horizontal stripes indicate S2 (attraction) and S6 (pre-attraction), black indicates S3

(Tracing), dark grey indicates S4 (transportation) and S5 (pre-transportation). Thedisabled robots are indicated by diagonal stripes.

The elapsed time until the robots begin each action is shown in Table 3. Eighty-nine seconds after the start of the trial, a robot found the food and tried to transport itto the nest (Fig. 6-A). Four robots were attracted at the same time by the pheromonetrail laid down by the first robot. As a result, four ¡five?¿ robots transported thefood, using pheromone communication. When the food began to move (Fig. 6-B),we stopped two of the robots from functioning so as to impede cooperative trans-portation. The five robots still performed repeated actions of attraction, tracing and


Table 3 Events of robustness experiment

event detail timeA finding food 89 [s]B starting transportation 309 [s]C task completed 2,990 [s]

transportation. At around 2,100 [s], all of the robots took part in transportation, andthey successfully brought the food to the nest at 2,990 [s] (Fig. 6-C). Our interven-tion did not affect the systematic function of the swarm.

The results of this experiment indicated that this swarm robot system is robustagainst internal variation. However, it should be noted that this robustness is de-rived from system redundancy, i.e., there needs to be more robots than required tocomplete the task.

4.3 Flexibility of the Swarm

To determine the flexibility of the swarm, we performed an experiment with sevenswarm robots implementing cooperative transportation using pheromone communi-cation. At 60 [s] after starting cooperative transportation, we changed the weight ofthe food from 3.58 [kg] to 5.28 [kg], transportation of which requires at least fourrobots. Figure 7, which has the same axis representations and colour patterns as Fig.6, shows the experimental results. Table 4 shows the elapsed time in the same wayas in Table 3. A robot found the food and began trying to transport it at 233 [s](Fig. 7-A). The robots laid down pheromone trails four times before the first coop-erative transportation by three robots at around 600 [s]; soon after this, four robotswere attracted at the same time (Fig. 7-B). Sixty seconds after smooth transporta-tion started, we increased the weight of the food at 762 [s] (Fig. 7-C). Transportationstopped due to a lack of sufficient number of participants. However, the robots soonlaid down the pheromone trail again at around 800 [s]. As a result of this, five robotsaggregated, and they began to move the food again at 923 [s] (Fig. 7-D). At 1,509[s], having transported the food to the nest, they had completed the task (Fig. 7-E).

Table 4 Events of flexibility experiment

event detail timeA finding food 233 [s]B 1st transportation start 702 [s]C changing food mass 762 [s]D 2nd transportation start 923 [s]E task completed 1,509 [s]

Changes in the external state did not cause any systematic problems for the swarmrobots, and they dealt with the changes in a concerted manner. Thus, our newlydeveloped robots possess both robustness and flexibility as a swarm.


Fig. 7 Experimental results of flexibility experiment

5 Conclusion

As mentioned in 4.1, pheromone communication contributes to a reduction in tasksolution time, especially when the density of the robots is relatively low as shownin Fig. 5. This suggests the effectiveness of pheromone communication. Even whenthere are only a limited number of swarm robots in a given environment, they cansolve the cooperative transportation task by making use of pheromone communica-tion. This means that the effectiveness of pheromone communication is dependenton the density of individuals in the environment.

As shown in Fig. 6, the robustness is likely mainly due to redundancy, which alsoprovides the swarm with flexibility (Fig. Fig. 7). When a swarm is not redundant,the swarm robots could solve a task, but they would still hardly shift smoothly toa new task, such as approaching another food. Our future studies on swarm robotswill aim to clarify two crucial interrelationships: i.e., those among simultaneousmultitask processing, the swarm behaviour and its redundancy, and those amongrobustness, flexibility and redundancy.

References

1. Fujisawa, R., Imamura, H., Hashimoto, T., Matsuno, F.: Development of multi-robotscommunicating by pheromone trail. Information Processing Society of Japan, Transac-tions on Mathematical Modeling and its Applications 2(2), 81–90 (2009) (in Japanese)

2. Bonabeau, E., Dorigo, M., Theraulaz, G.: Swarm Intelligence. Oxford University Press(1999)

3. Wilson, E.O.: Sociobiology: The New Synthesis. Belknap Press of Harvard UniversityPress, Cambridge (1975)


4. Matsuka, M., Kitano, H., Matsumoto, T., Oono, M., Gokan, N.: Biology of Insect. Tam-agawa University Publishing (1992) (in Japanese)

5. Katsumata, A., Ozaki, M.: Chemical communications in ants. Journal of ComparativePhysiology and Biochemistry 24(1), 3–17 (2007)

6. Karlson, P., Butenandt, A.: Pheromones (ectohormones) in insects. Annual Review ofEntomology 4, 39–58 (1959)

7. Ito, K. (ed.): Emergence of Intelligence. NTT Publishing (2000) (in Japanese)8. Sugawara, K., Kazama, T., Watanabe, T.: Foraging behavior of interacting robots with

virtual pheromone. In: Proc. IEEE/RSJ Int. Conf. on Intelligent Robots and Systems(IROS 2004), vol. 3, pp. 3074–3079 (2004)

9. Garnier, S., Tache, F., Combe, M., Grimal, A., Theraulaz, G.: Alice in pheromone land:An experimental setup for the study of ant-like robots. In: Proc. IEEE Swarm IntelligenceSymposium (SIS 2007), April 1-5, pp. 37–44 (2007)

10. Shimoyama, I., Kanzaki, R.: Biological type micromachine. Journal of Society ofBiomechanisms 22(4), 152–157 (1998) (in Japanese)

11. Purnamadjaja, A.H., Russell, R.A.: Guiding robots behaviors using pheromone commu-nication. Autonomous Robots 23(2), 113–130 (2007)

12. Dorigo, M., Tuci, E., Groß, R., Trianni, V., Labella, T.H., Nouyan, S., Ampatzis, C.,Deneubourg, J.-L., Baldassarre, G., Nolfi, S., Mondada, F., Floreano, D., Gambardella,L.M.: The SWARM-BOTS Project. In: Sahin, E., Spears, W.M. (eds.) Swarm Robotics2004. LNCS, vol. 3342, pp. 31–44. Springer, Heidelberg (2005)

13. Mondada, F., Bonani, M., Guignard, A., Magnenat, S., Studer, C., Floreano, D.: Super-linear Physical Performances in a SWARM-BOT. In: Capcarrere, M.S., Freitas, A.A.,Bentley, P.J., Johnson, C.G., Timmis, J. (eds.) ECAL 2005. LNCS (LNAI), vol. 3630,pp. 282–291. Springer, Heidelberg (2005)

14. Kube, C.R., Bonabeau, E.: Cooperative transport by ants and robots. Robotics and Au-tonomous Systems 30(1-2), 85–101 (2000)

15. Fujisawa, R., Imamura, H., Hashimoto, T., Matsuno, F.: Communication UsingPheromone Field for Multiple Robots. In: Proc. IEEE/RSJ 2008 International Confer-ence on Intelligent Robots and Systems (2008)

16. Fujisawa, R., Dobata, S., Kubota, D., Imamura, H., Matsuno, F.: Dependency by Con-centration of Pheromone Trail for Multiple Robots. In: Dorigo, M., Birattari, M., Blum,C., Clerc, M., Stutzle, T., Winfield, A.F.T. (eds.) ANTS 2008. LNCS, vol. 5217, pp.283–290. Springer, Heidelberg (2008)

17. Fujisawa, R., Imamura, H., Hashimoto, T., Matsuno, F.: Swarm Intelligence of Multi-Robot Using Pheromone Trail. In: Proc. The 2nd Internal Symposium on Mobiligence,pp. 203–206 (2007)

[springer tracts in advanced robotics] distributed autonomous robotic systems volume 83 ||...

Documents