Indoor Map-Constrained Robot Localization Using Incidental Radio Observations Joydeep Biswas The Robotics Institute Carnegie Mellon University Pittsburgh PA 15213 Email: joydeepb@cs.cmu.edu Manuela Veloso Computer Science Department Carnegie Mellon University Pittsburgh PA 15213 Email: veloso@cs.cmu.edu Abstract—We are interested in equipping an indoor mobile robot with a robust localization algorithm. We assume that the robot is performing some task that requires it to continuously move, and localization needs to be incidentally achieved. A few other known efforts, e.g., a museum-guide robot, a service robot, or a soccer robot, have addressed similar indoor inci- dental localization and navigation problems, but have mostly relied on visual-based landmark perception. Such tasks can include frequent stoppages that can play an important role in reducing localization uncertainty. In this work, we assume that the environment includes static radio nodes, which can be queried for localization landmark information. We introduce a representation of the indoor map in terms of both rigid map constraints, e.g., corridor width and connectedness, and as a set of sampled location points with signal strengths defining the a priori map knowledge. The dynamic radio signal strength from the sensor network nodes, along with the discrete map points is used to generate a continuous localization probability distribution. This probability distribution, combined with the motion data of the robot and the rigid map constraints reults in a novel indoor localization algorithm. Two novel approaches in our algorithm compared to other algorithms are the method of generation of an estimate of the perceptual model that is continuous in space, and the use of rigid map constraints for weighting location beliefs. We demonstrate the algorithm in an autonomously moving robot that interacts with a real sensor network in our building. I. INTRODUCTION An autonomous robot needs to be location aware in order to meaningfully execute its tasks. Numerous approaches address the localization problem with varying degrees of success. Most robots operating outdoors can tolerate localization errors on the order of one meter, generally using GPS readings. Very few devices can consistently receive GPS signals indoors with appropriate accuracy. A variety of approaches to indoor localization have been developed that rely on the detection of features by the robot’s sensors, and estimation of the range and orientation with respect to known landmarks. The landmarks chosen can be artificial, like colour-coded markers, or pre-existing objects like doorways and light fixtures. Some indoor localization algorithms use optical sensors including cameras, laser scanners, and infrared sensors, but such approaches do not perform as well in crowded indoor environments, where people, robots and other mobile obsta- cles may obstruct the view of the sensors. Algorithms that attempt to perform Simultaneous Lo- calization and Mapping (SLAM) commonly use particle filters for representing the location hypotheses of the robot. However, the hypotheses are permitted to evolve irrespective of map based constraints. In this work, we use radio frequency (RF) beacons, as well as wireless networks to provide indicative range information using received signal strength. This approach is not affected by changes in the visual appearance of the environment, although it is still susceptible to local obstacles like people and other robots. Our approach at tackling the localization problem is mo- tivated by the development of an indoor visitor companion robot. The environment in which the robot is to operate is an office building with a series of interconnected hallways, which can be traversed by the robot. As in any busy office building, the hallways have people and other robots involved in their own tasks. There will also be obstructions like carts, tables, chairs and boxes left in the hallway temporarily. The robot should be free to move about, performing its own tasks without having to stop to localize itself. This implies that any sensory inputs that assist in the localization of the robot should be incidental, without requiring the robot to stop for readings just to obtain a localization estimate. In this paper, we describe an algorithm based on Markov localization for localizing a robot on a known map using RF signal strength from the nodes of a sensor network. In the derivation of the sensor prediction model, we assume no knowledge of the propagation characteristics of the RF signals. Our algorithm for localization of the robot consists of two stages. In the first stage, the Map Learning Phase, we introduce a representation of the map using a set of vertices V , and a set of edges E connecting them. The vertices are distributed in space according to a given distribution, e.g. uniformly at 1m intervals. Fig. 1 shows an example of the graph representation of a section of a map. In the figure, the vertices v 1 , v 2 , v 3 and v 4 are shown connected by edges (dashed lines). Three RF nodes, n 1 , n 2 and n 3 are also shown. To complete the Map Learning Phase, the robot is manually driven around, and at every vertex, the robot is