Swarm-bots project

All-terrain navigation requires the ability to cope with a generic rough terrain, comprising slopes and obstacles that have to be climbed. The swarm-bot has been designed intentionally for this purpose, as it presents mechanical features that allow a single individual to cope with moderately rough terrains. In this section, we will show how cooperation among individual s-bots is beneficial for rough terrain navigation, whenever the individual abilities are too constraining.

In order to assess the ability of a swarm-bot in rough terrain navigation, we performed two tests using the controllers evolved for coordinated motion (see here). In the first test, a group of s-bots is placed in an arena with hills, shown in Figure 1 left. The task of the robots is to move as far as possible from the initial position notwithstanding the presence of the hills. This scenario has been used to evaluate if the controllers evolved with the linear s-bot formation on a flat terrain can still work in an arena with hills. The test is particularly important since when an s-bot moves up the hill, the turret leans down hill and tends to generate a traction that might make the s-bot turn.

In the second test, we used the same controllers to drive a group of s-bots in a complex environment with a very rough terrain (see Figure 1 right). In this case, the controllers evolved for coordinated motion using the simple simulation model are tested using the detailed one. Here, the generalisation capabilities of the evolved controller must prove efficient not only with respect of new environmental conditions, but also with respect to the physical embodiment of the robots.

Figure 1. Left: A group of four assembled s-bots with flexible links moves in an arena with hills. Right: A group of s-bots simulate using the detailed model performs coordinated movements on rough terrain.

In the first set of experiments, the s-bots and controllers used are those described here. In particular, the neural controllers used here are the same evolved with four s-bots forming a linear swarm-bot for coordinated motion in a flat terrain. As illustrated in Figure 1 left, there are randomly scattered hills in the arena. The hills have been implemented as 40 cm radius spheres emerging from the ground for some centimetres. In particular, 5 different tests have been run with hills' heights measuring from 0 to 4cm.

In the second set of experiments, the controllers drive a group of s-bots simulated using a very detailed model of the real robot, as shown in Figure 1 right. The environment is a rough terrain simulated as an height field. The swarm-bot is placed in the centre of the arena and has to move as far as possible within the limited amount of time available in each trial, without changing direction or remaining trapped into steep concave regions. We performed some tests varying the maximum height of the terrain in the range [0,20] cm.

The neural networks evolved to control swarm-bots in a flat arena are able to deal with a rough terrain with hills of some height. While single s-bots often fall over when they encounter the hills, swarm-bots succeed to pass them. We performed a number of trials for each experimental condition, obtained varying the height of the hills scattered in the environment. The results show that the fitness starts to decrease sensibly when the hills are above 2 cm high. However, up to 4 centimetres the performance is still acceptable (above this height the swarm-bot has difficulties in finding the initial coordination if positioned in a valley). Another series of test confirmed that a swarm-bot outperforms the single s-bot. Direct observation of the behaviour reveals that the single s-bot often overturns when it encounters the hills, while this never happens to the swarm-bot.

The above tests confirm that a swarm-bot is more efficient than the single robots in navigation on rough terrain. They also show that the controllers developed for the coordinated motion task easily generalise to new environmental conditions. Therefore, even changing the basic interactions among robots and between robot and environment, the coordinated motion behaviour is still efficient. The above experiments were conducted using a simple simulation model of the s-bots, which differ consistently from the real hardware above all in the traction system. When dealing with rough terrain conditions, the interactions between the s-bot's traction system and the ground should be carefully simulated, as they could deeply influence the effectiveness of the developed behaviour. For this reasons we devised the scenario depicted in Figure 1 right, where the s-bots and the rough terrain are carefully simulated. We tested the coordinated motion behaviour in these conditions, varying the roughness of the terrain. We performed a number of evaluation trials in each experimental condition, and looked at the fitness of a linear swarm-bot formation composed of 4 s-bots. Also in this case, the fitness is proportional to the distance covered by the group. The obtained results show that there is no loss in performance when using the detailed simulation model, even with a maximum height of the terrain of 7.5cm. For increasing roughness, the fitness slightly decreases, but the performance drop is not as apparent as in the previous tests. This result is principally due to the simulation model used in these experiment, which, closely simulating the treels system of the s-bot, is better suited for rough terrain navigation.

Moving on rough terrain

Experimental setup

Results