Centibots: Very Large Scale Distributed Robotic Teams

ly, the zippering process lets us take any partial maps produced by any robots and put them together, once a common location (colocation) between their trajectories has been identified (note that colocation is transitive). In order to determine these common locations, we developed an efficient algorithm that sequentially estimates the relative locations between robot pairs as they explore an environment [6]. The approach considers only pairs of robots since the complexity of estimating map matches is exponential in the number of robots considered jointly. For each robot pair, the technique uses an adapted particle filter to estimate the position of one robot in the other robot’s partial map. By estimating the posterior over robot positions both inside and outside the partial map, the approach is also able to estimate whether or not there is an overlap between the robots’ maps. To accurately determine the overlap probability, we developed a hierarchical Bayesian technique that learns a prior over the structure of indoor environments and uses the structural model to estimate the certainty of map matches [2, 10]. Active Multi-robot Colocation and Exploration Virtually any map matching technique can generate false-positive matches, especially in large, highly symmetric environments. A wrong map match between two robots can generate subsequent wrong map matches with other robots. Thus, undoing a wrong match requires considering all other map matches as well. To avoid the complexity resulting from wrong matches, we developed a technique that coordinates robots to actively verify whether or not a map match hypothesis is correct. The approach is integrated into a decision-theoretic multi-robot exploration strategy (see [6] for details). Figure 5(b) shows an example run using our coordination technique. The two robots, A and B, start from different, unknown locations. Initially, the robots explore on their own. As they explore, each robot estimates the other robot’s location in its own map, using the modified particle filter mentioned above. When deciding where to move next, both A and B consider whether it is better to move to an unexplored area (frontier), or to verify a hypothesis for the other robot’s location. At one point, B decides to verify a hypothesis for A’s location. It sends A the message to stop and moves to A’s hypothesized location. Upon reaching this location, both robots check the presence of the other robot using their laser range-finders (robots are tagged with highly reflective tape). When they detect each other, their maps are merged using the zippering process described above. From then on, they explore the environment in a coordinated way. If a hypothesis verification fails, on the other hand, then the hypothesis is simply deleted, and all robots keep on exploring. Our coordination technique works for more than two robots. Multiple robots can share a common map and coordinate to explore and verify hypotheses for the locations of other robots. Since each map merge operation increases the number of robots sharing a common map, team coordination improves over time. Howard et al. [5] also use robot detections to merge maps (and close loops). In contrast to our active colocation technique, their approach is purely passive in that robots have to detect each other coincidentally. Passive map merging can result in significant delays, for example, when one robot follows the path of the other robot and never actually detects it. Exploration with Limited Communication Robots form so-called exploration clusters, which are groups of robots that share a common map. A team leader robot uses this map to coordinate the other robots. New robots can be added to a cluster, once their relative location with respect to the cluster map is determined. Maps are represented compactly as sets of laser range-scans annotated with robot poses and probabilistic links (scans are recorded only every 50cm). Each robot integrates its observations into its own map, and broadcasts the information to the other robots. While most of the other robots only store this data, the team leader integrates all the sensor information it receives. Thus the team leader has a complete and consistent map representing the data collected by all robots in the cluster. Frequently, this map is broadcast to the other robots, in order to guarantee consistency. The data can be (a) Start robot B Start robot A