Dataset
The dataset consists of 17 distance sensors recorded during navigation of a mobile robot in a simulated environment called SINAR.
SINAR is a 2D autonomous robot simulator introduced in [1], where the mobile robot interacts with the environment by distance sensors and by one actuator which controls the movement direction (turning).
The robot model has 17 distance sensors distributed uniformly over the front of the robot, from -90◦ to +90◦. The sensors are limited in range such that they saturate for distances greater than 300 distance units (d.u.), and exhibit Gaussian noise N(0, 0.01) on their readings. A value of 0 means near some object and a value of 1 means far or nothing detected. At each iteration the robot is able to execute a direction adjustment to the left or to the right in the range [0, 15] degrees and the speed is constant (0.28 distance units (d.u.)/s). The SINAR controller, described in [1], is a reactive intelligent navigation system made of hierarchical neural networks which learn by interaction with the environment. After learning, the robot is able to efficiently navigate and explore environments, during which the input signal is built by recording the 17 distance sensors of the robot.
Environment
The environment is a big maze with 64 predefined locations spread evenly around the environment (represented by small labeled triangles). The input signal (17 distance sensors) is generated while the robot is navigating in this environment for 350.000 timesteps.
Error Metrics
If learning is unsupervised, we use the place cell reconstruction method from [2] to estimate the most likely robot position given the activation of place cells, after training.
The error is the Euclidean distance between the true robot position and the predicted robot position.
References
[1] E. A. Antonelo, A.-J. Baerlvedt, T. Rognvaldsson, and M. Figueiredo, “Modular neural network and classical reinforcement learning for autonomous robot navigation: Inhibiting undesirable behaviors,” in Prceedings of the International Joint Conference on Neural Networks (IJCNN), Vancouver, Canada, 2006, pp. 498– 505
[2] K. Zhang, I. Ginzburg, B. L. McNaughton, and T. J. Sejnowski, “Interpreting neuronal population activity by reconstruction: unified framework with application to hippocampal place cells.” J Neurophysiol, vol. 79, no. 2, pp. 1017-1044, Feb 1998.
