Control >> Cooperative transport >> Two minimalist robots

Transport by a team of two minimalist robots

This section addresses the problem of cooperative transport of a prey by a team of two simple s-bots that are capable of establishing physical connections between each other and with the prey.

In this study, the s-bots are not provided with a rotational degree of freedom between the turret and the chassis (i.e., the orientation is locked). There is no explicit communication among s-bots. Since the s-bots cannot sense each other, interactions are restricted to indirect interactions via the environment which consists of the prey and a light-emitting beacon, or interactions through the s-bot's embodiment.

The task is to control a group of two simple s-bots so that they cooperate to transport a heavy prey as far as possible within a fixed time period (see Figure 1). The direction of movement of the prey is not predetermined. Since the prey cannot be moved by a single s-bot, coordination among both s-bots is necessary.

Figure 1. The Task.

Experimental setup

The experiment is carried out using a 3D physics simulator. The simulated environment consists of a flat ground, a light-emitting beacon, and a prey. The prey is modeled as a cylinder of mass 500g.

The s-bot can control its left and right wheels and the gripper element. The gripper element is modeled as controllable, sticky box heading forward. If the element is in gripping mode, a connection will be established as soon as the gripper element is in contact with a grippable object. Once the gripper is set to the open mode, any gripped object will be released.

The s-bot is equipped with an omni-directional camera and a gripper status sensor. A summary of the information provided by each sensor is presented in the following table.

The s-bots are controlled by simple recurrent neural networks that are shaped by an evolutionary algorithm. All the s-bots of a group transporting a prey are initially equipped with an identical neural network. The fitness function rewards the group solely for high-level objectives (e.g. maximize the distance the prey is moved). Therefore, the fitness function does not restrict (explicitly) the type of behavior by which these high-level goals should be reached.

Results

The experimental setup described above has been used in ten independent evolutionary runs of 150 generations each. Overall, the controlled groups of s-bots act quite robustly with respect to various kinds of noise concerning the sensors and the actuators and with respect to different initial placements in the environment. Evolved controllers attained a satisfactory level of performance. They succeeded in controlling larger groups of robots to transport bigger prey. However, we discovered that the performance is very sensitive to the prey's characteristics (shape and dimension).

Example movies:
• Solution that makes no use of gripper element (MOV, 1.8 MB).
• Solution that makes use of gripper element (MOV, 1.7 MB).
• Example with five robots transporting a big prey (without use of gripper element) (MOV, 2.2 MB)
  
• Example with five robots transporting a big prey (with use of gripper element) (MOV, 2 MB)
  

References

• Groß R. and Dorigo M. Evolving a Cooperative Transport Behavior for Two Simple Robots, In Liardet P., Collet P., Fonlupt C., Lutton E., and Schoenauer M., editors, Artificial Evolution - 6th International Conference (EA 2003), volume 2936 of Lecture Notes in Computer Science, Springer Verlag, Berlin, Germany (2004) 305-317

Control >> Cooperative transport >> Two minimalist robots