Learning local behavioral sequences to better infer non-local properties in real multi-robot systems
By Taeyeong Choi, Sehyeok Kang, and Theodore P. Pavlic
We attach Thy-ReTLo, a dataset specialized for Remote Teammate Localization (ReTLo) on a commercial, two-wheeled robotic platform, Thymio [2], where a robot in a multi-robot team only uses observation about its nearest neighbor to infer the locations of all other far teammates. The dataset provides global coordinates and orientation of each robot, collected from real Thymio robots. The recording occured twice per second by a central computer monitoring and detecting the change of locations through a overhead camera. We offer two separate cases based on the team size: 3-robot team and 5-robot team.
All teams bulid upon the motion rules defined in [1] according to the role of robot. We encourage to train an inference model only in the 3-robot team and execute it not only in the same size but also in the 5-robot case. Thus, we utilize 60% and 10% of data from 3-robot team as Training and Validation set, respectively. The rest of 20% from 3-robot team and all from the 5-robot are used to test the finally obtained model.
More details about the proportions are shown below:
A commercial, two-wheeled robotic platform, Thymio, is used for all data collection. Each robot is powered by a Raspberry Pi, which communicates with a central monitoring computer via an overhead camera to simulate a better proximity sensor and a virtual GPS system. Through the communication, the Thymio's can determine the next motion.
Also, an arena of size 2.5m x 1.9m is used across all types of recordings, on which all robots covered by cardboards, for better detection, are programmed to move, as shown below:
The csv data from k-robot team is located in ./Datasets/kRobots. Each row contains an instance of pose information for all robots, recorded in global coordinates for 26 sequenctial time steps (13 seconds). t_r_x, t_r_y, and t_r_o are the x,y-coordinates and orientation of robot r at timestep t where H, T, and Fk represent Head, Tail, and Follower k where Follower 1 is the nearest one to Head. Since the recording was performed such that the instances have at least 7-second time gap among them, it can be said that there is little motional dependency between any two instances. The number of instances for each case are described below:
If you have any questions on use of this data, please send an email to tchoi4@asu.edu.
[1] Taeyeong Choi, Theodore P. Pavlic, Andrea W. Richa, Automated synthesis of scalable algorithms for inferring non-local properties to assist in multi-robot teaming, 13th IEEE Conference on Automation Science and Engineering (CASE), 2017.
[3] https://www.raspberrypi.org/products/raspberry-pi-3-model-b/