João Macedo

Locating odour sources with mobile robots is a difficult task with important applications. The most relevant environmental variables necessary to carry-out this task are the odour concentration and the wind velocity. This paper addresses the design and development of 2D thermal anemometers for very low wind speeds, such as the ones found indoors, and validate their performance with a mobile robot moving in a large wind tunnel. These anemometers use the wind disturbance around a prismatic shape to estimate the orientation of the airflow. Common shapes, such as cylinders and other regular prisms, were studied with a computational fluid dynamics (CFD) tool and with experimental prototypes. The prototypes were instrumented with a set of small self-heated thermistors, placed around the prism, in the measuring points of interest. This work employed Random Forests to process the output of the thermal elements and classify the airflow orientation in 36 intervals, producing anemometers with 10° resolution. Although all of the studied shapes could be employed to reconstruct the wind velocity vector, the one with best performance in terms of smoothness of response and robustness to turbulent data is a cylinder, which suggests that this method can be latter adapted to common cylindrical robots, exploiting their body as the structure of a similar anemometer.

Detecting and locating odour sources with mobile robots is hard, yet an interesting problem for many real world applications. In nature, this problem is daily and successfully addressed as a way to survive. This success motivated researchers to adapt the observed biological search strategies for robots. Each of these strategies is meant to operate under specific environmental conditions and, despite being inherently different, they can be decomposed into a small set of behaviours. The present paper compares the performance of those behaviours in finding and tracking odour plumes to their sources, under diverse environmental conditions. The experimental results show that the performance of each behaviour is highly dependent on its parameters, as well as on the involving environmental conditions and thus, the behaviour employed must be carefully selected. The resulting knowledge may be used by the community for producing better performing strategies, either by hand-designing them or through learning methods.

This paper presents a geometric crossover operator for Tree-Based Genetic Programming that acts on the syntactic space, where each expression tree is represented in prefix notation. The proposed operator is compared to the standard subtree crossover on a symbolic regression problem, on the Santa Fe Ant Trail and on a classification problem. Statistically validated results show that the individuals produced using this method are significantly smaller than those produced by the subtree crossover, and have similar or better performance in the target tasks.

Using robots to locate odour sources is an interesting problem with important applications. Many researchers have drawn inspiration from nature to produce robotic methods, whilst others have attempted to automatically create search strategies with Artificial Intelligence techniques. This paper extends Geometric Syntactic Genetic Programming and applies it to automatically produce robotic controllers in the form of behaviour trees. The modification proposed enables Geometric Syntactic Genetic Programming to evolve trees containing multiple symbols per node. The behaviour trees produced by this algorithm are compared to those evolved by a standard Genetic Programming algorithm and to two bio-inspired strategies from the literature, both in simulation and in the real world. The statistically validated results show that the Geometric Syntactic Genetic Programming algorithm is able to produce behaviour trees that outperform the bio-inspired strategies, while being significantly smaller t...

This paper addresses the problem of controlling a group of mobile robots to track an odour plume to its source. To perform this task in real environments, it is important that the robots are able to adapt to a changing world, and use the experience gained to improve their performance. We address this task with Genetic Programming to evolve the controllers for the robots. Two evolutionary approaches are proposed and compared to a variant of the Silkworm Moth algorithm, that has been modified to take advantage of multi robot systems. The statistically validated results showed that, in the groups of robots where significant differences were found, the evolved controllers were able to find the odour plume faster and converge to its source better than the Silkworm Moth approach.

This paper presents a mobile robot to detect indoor smoldering fires. The robot uses a custom multisensory system able to measure a set of environmental parameters including CO, CO2, NO2, O3 and airborne particles. The paper proposes a sensory fusion method to detect fires with this system and evaluates that method in realistic environments. The obtained results show that the proposed method is able to detect fires which are undetected by a commercial system and that it is able to discriminate between burning materials.

Locating odour sources is a hard task that has been addressed with a large variety of AI methods to produce search strategies with different levels of efficiency and robustness. However, it is still not clear how to evaluate those strategies. Simply evaluating the robot's ability to reach the goal may produce deceptive fitness values, favouring poor strategies that do not generalise. Conversely, including prior knowledge may bias the learning process. This work studies the impact of evaluation functions with various degrees of prior knowledge, in evolving search strategies. The baseline is set by performing multiple evaluations of each strategy with a function that only evaluates the task efficiency. A function was found that is able to produce strategies with equivalent performance to those of the baseline, whilst performing a single evaluation.

The estimation of the parameters of an odour source is of high relevance for multiple applications, but it can be a slow and error prone process. This work proposes a fast particle filter-based method for source term estimation with a mobile robot. Two strategies are implemented in order to reduce the computational cost of the filter and increase its accuracy: firstly, the sampling process is adapted by the mobile robot in order to optimise the quality of the data provided to the estimation process; secondly, the filter is initialised only after collecting preliminary data that allow limiting the solution space and use a shorter number of particles than it would be normally necessary. The method assumes a Gaussian plume model for odour dispersion. This models average odour concentrations, but the particle filter was proved adequate to fit instantaneous concentration measurements to that model, while the environment was being sampled. The method was validated in an obstacle free cont...

Locating odour sources with robots is an interesting problem with many important real-world applications. In the past years, the robotics community has adapted several bio-inspired strategies to search for odour sources in a variety of environments. This work studies and compares some of the most common strategies from a behavioural perspective with the aim of knowing: (1) how different are the behaviours exhibited by the strategies for the same perceptual state; and (2) which are the most consensual actions for each perceptual state in each environment. The first step of this analysis consists of clustering the perceptual states, and building histograms of the actions taken for each cluster. In case of (1), a histogram is made for each strategy separately, whereas for (2), a single histogram containing the actions of all strategies is produced for each cluster of states. Finally, statistical hypotheses tests are used to find the statistically significant differences between the beh...

The tasks of odor detection, plume tracking and odor source localization constitute an important, yet complex, real world problem. One possible solution for them is based on the use of a group of mobile robots whose controllers have to be defined. Artificial Neural Networks (ANN) have already been used as controllers, but the task of hand defining their topology and parameters can be very challenging and time consuming. In this paper, we propose an approach to evolve, rather than design, ANN-based controllers. Our approach relies on Genetic Programming (GP), a family of stochastic search procedures loosely inspired by the biological principles of Natural Selection and Genetics. We compare our approach with a classic one, inspired by the chemotaxis behavior of the E. coli bacteria. Our results show that this approach is able to outperform the chemotaxis in the experiments performed.