Two modules learning to locomote by self-reconfiguration in a 2D simulated world: the learning episodes on video correspond to the episodes between the two vertical lines on the reward graph.
Nine modules moving East by self-reconfiguration in a 2D simulated world: they are executing a policy learned by the LLGAPS algorithm, which can be seen on the probability graph.
The video clips above demonstrate a distributed coverage control algorithm. The robots are meant to maintain coverage of the whole region while collecting in greater concentration over areas where the light is brighter. Robots in the left clip have been adjusted to be less sensitive to the light than those in the right clip.
The video clips below demonstrate dynamic task assignment on a small group of robots. The goal distribution is specified in terms of the percentage of red, green, and blue robots. This goal distribution is relayed to the swarm at run-time. All of the algorithms are robust to the removal or addition of robots at any time, which is critical for implementation in any real multi-robot system.
We considered four algorithms: Algorithm Random-Choice selects tasks randomly, but runs in constant time. Algorithm Extreme-Comm compiles a complete inventory of all the robots on every robot, runs quickly, but uses a great deal of communication. The Card-Dealer's algorithm assigns tasks to individual robots sequentially, using minimal communications but a great deal of time. The Tree-Recolor algorithm is a compromise between Extreme-Comm and Card-Dealer's, balancing communications use and running time.
Algorithm Random-Choice is the simplest solution: robots choose a given task with probability equal to the relative size of that task subgroup in the target distribution. This algorithm requires no communication and completes immediately. However, there is a high probability that it will fail to achieve the target distribution in small to medium-sized swarms (10-50 robots).
In the Extreme-Comm Algorithm, each robot uses local interactions to build a complete list of all other robots in the swarm, and uses this list to determine its task. Though fast and accurate, the algorithm requires a large amount of inter-robot communications, and is impractical in all but the smallest groups.
The Card-Dealer's algorithm uses minimal communications by sequentializing the task-assignment problem into a series of stages. The robots "count-off" to allocate tasks, and the algorithm requires a long execution time.
Finally, the Tree-Recolor algorithm is a compromise between Extreme-Comm and Card-Dealer's, balancing communications use and running time by using gradient trees.
Research sponsored by Boeing