Control >> Cooperative transport >> Blind and non-blind s-bots

Transport by pre-attached blind and non-blind s-bots

This section details the experimental work concerning the transport module that is used to control s-bots which are not capable of localizing the target location.

Controller

  1 repeat
  2 (i_1, i_2, i_3, i_4) = getTractionSensorReadings()
  3 i_5 = detectStagnation(torque on tracks and turret)
  4 i_6 = orientation of chassis
  5 (o_1, o_2) = neuralNetwork(i_1, i_2, i_3, i_4, i_5, i_6)
  6 α = (o_2 + 2 * i_6) / 3
  7 if (torque problem on turret or tracks) then
  8 execute recovery move
  9 else
10 setTurretOrientation(α)
11 if (soft alignment sufficient) then
12 move and softly re-align towards α (maxspeed o_1)
13 else
14 turn on the spot towards target
15 endif
16 endif
17 until (timeout reached)

Algorithm 1: Transport module to control blind s-bots.

Algorithm 1 describes the control module for the transport by blind s-bots. The basic concept of the control is identical to the one studied in simulation: a neural network controller determines the orientation of the chassis as well as the transport speed for the tracks. We derived the controller by modification from the controller for the transport by non-blind s-bots (see here).

The neural network is a simple recurrent neural network with 4 hidden nodes. The activations of the input nodes i_1, i_2, i_3 and i_4 represent the tractional force between the s-bots turret and the chassis concerning four different directions (front, left, back, and right). The activation of input node i_5 is 1 iff a problem concerning high torque on the wheels or the turret has occured during the last 4 control steps. Thus, i_5 indicates potential stagnation. i_6 corresponds to the orientation of the chassis with respect to the turret. The activations of all input nodes are scaled to the range [0,1].

Output o_1 specifies an upper limit for the speed the s-bot can move with during transport (line 12). Output o_2 determines the desired direction of the chassis.

Experimental Setup

In this study only Environment A is considered. The basic setup has been retained unchanged. The s-bots start being attached to the prey, either directly or within a chain formation. They are supposed to transport the prey within a fixed time period as far as possible towards the target location. Since some additional time might be required for a blind s-bot to adapt its behavior to the rest of the group, we extended the period of a trial by 5 seconds (in total 20 seconds).

Those s-bots of the group that are non-blind, are controlled by the control module for the non-blind s-bots (see here).

Figure 1: Spatial arrangement of 2-3 s-bots around the prey (gray). The blind s-bot is shown in black, the other in white. The transport target (not shown) is a light source, 250cm to the right of the prey.

Figure 2: Spatial arrangement B3 of two non-blind (the two on the left) and one blind s-bots (the one on the right).

In this study, 2 to 3 s-bots are used, one of which is blind. The number of different spatial arrangements has been limited to the discrete set illustrated in Figure 1. For instance, Figure 2 shows the spatial arrangement B3. All arrangements fulfill the following condition that each non-blind s-bot is initially arranged in such a way that no obstacle (prey or teammates) is shadowing the target location.

For the purpose of demonstration, in some trials the blind s-bot has been equipped with a paper belt which covers the target sensors (proximity sensors in ambient light mode) - thus the s-bot cannot perceive the light anymore (see Figure 2). We removed the belt in some trials to allow for collecting data for further analysis.

Results

Figure 3: Transport performance. 2 non-blind (hand-coded controller described previously and 1 blind s-bot (neural network controller described in this section). Six observations per box.

Figure 3 shows the distance gained within transport of the prey by three s-bots of which one is blind. It is worth noting that the cooperation of at least two s-bots is necessary to move the prey. By looking at this plot, we cannot conclude that the blind s-bot contributes to the group performance. Indeed, this would be a surprise, since the prey is so light, that even with a hand-coded controller and having no blind s-bots, there seems to be no benefit by using three robots (as seen previously). In configuration B0 the blind s-bot behaves disruptive, in the other three cases the performance is much better. The question arises how good the performance would be, if the blind s-bot would not move its tracks. We tested this for each arrangement once. Every time, the action of the s-bot disrupted the performance of the group drastically. The corresponding distance gain values are 3, 18, 13, and 39cm.

Figure 4: Transport performance. 1 non-blind (hand-coded controller) and 1 blind s-bot (neural network controller). Only 3 observations per box.

We consider a team of one non-blind s-bot and a blind one. We have seen that one s-bot by itself is nearly incapable of transporting the prey. In simulation, we have shown that in the specific case of a team of one blind and one non-blind s-bot, the strategy to pull in a random direction can already lead to a positive distance gain (on average). For the 6 different arrangements that we tried once each, the observed distance gain is 11, 5, 78, -10, 145, and -8. Although the number of observations is very low, the outcome does support the idea that a) if the s-bot is pulling by chance in a good direction the prey gets closer to the target, but if the s-bot is pulling in a bad direction, the two s-bots are likely to cancel out each other.

If we control the blind s-bot by the evolved neural network instead, the performance seems to be very good (see Figure 4) and also the fluctuations are very low. The number of observations in this particular plot (18 in total) is too low for making conclusions about specific arrangements.

Overall, we tested a variety of spatial arrangement of two s-bots with the prey, and in 17 out of 18 cases, the blind s-bot could contribute to the groups performance. In 15 cases the prey could be moved for about 200cm, the performance fluctuations have been astonishingly low.

Example movies:

• Cooperative Transport by Blind and Non-Blind s-bots (Quicktime, 2.9 MB).

References

• Groß R. and Dorigo M. Group Transport of an Object to a Target that Only Some Group Members May Sense, In Yao X., Burke E., Lozano J. A., Smith J., Merelo-Guervós J. J ., Bullinaria J. A., Rowe J., Tiňo P., Kabán A., and Schwefel H.-P., editors, Parallel Problem Solving from Nature - 8th International Conference, PPSN VIII, volume 3242 of Lecture Notes in Computer Science, pages 852-861. Springer Verlag, Berlin, Germany, 2004