HTP Graphics

RAIN PROGRESS // The approach for exploration of a-priori unknown environments integrating Gaussian process models with Markov decision process planning has been published at the 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). For the publication, we tested on a real dataset of gamma radiation intensity observations collected by a Clearpath Jackal mobile robot at the University of Lancaster Neutron Laboratory. We demonstrated that our approach enables the robot to explore safely and efficiently: First, we showed that our GP model of radiation can achieve a level of prediction similar to the real dataset using only approximately one third of the datapoints (appropriately chosen by our algorithm). Then, in a set of realistic simulations obtained from the real dataset, we showed that, on average, our proposed algorithm is able to determine the radiation level at all reachable safe states (to within a defined standard deviation of 3 counts per second) in 52% of the time compared to a state-of-the-art algorithm, while carrying out 56% fewer observations. REMOTE INSPECTION UNIQUENESS // The approach we developed focuses on integrating Gaussian process (GP) models of radiation with a Markov Decision Process (MDP) model for decision-making. GPs allow for maintaining predictive models of radiation across the environment, updating them online with the observed values of radiation. Crucially, they can predict the value of radiation in each location of the environment, by modelling it as a Gaussian distribution. This allows the system to reason about its certainty over the value of radiation in a certain location, thus incorporating observations into a probabilistic planning framework in a principled Bayesian way. It extends state-of-the-art approaches by explicitly considering the uncertainty over radiation levels given by the GP in the transition function of the MDP used for planning. This allows for the robot to more safely and efficiently explore the environment. FUTURE ASPIRATIONS // We plan to implement this work in a Clearpath Jackal robot and demonstrate the radiation-aware safe exploration framework using University of Manchester's ThermoFisher RadEye sensors. Furthermore, we plan to generalise the framework in two ways. First, whilst the approach can safely explore and learn a predictive model of radiation in the environment, it assumes a map of the workspace is known a priori. To relax this assumption, we intend to integrate our approach with a simultaneous localisation and mapping (SLAM) approach, thus allowing for the safe construction of both a metric map and a radiation map of the environment. We will implement a smart radiation mapping system developed in the context of a collaboration between ORI and Createc Technologies, consisting of a Clearpath Jackal robot equipped with a CZT radiation detector and an integrated SLAM system developed at ORI. Second, the approach only allows for pure exploration and does not consider goal-driven behaviour (e.g.safely navigating to a specific to be inspected as quickly as possible). We will extend our approach to do so, thus allowing for monitoring systems that are also goal- driven, rather than the current completely exploration- driven approach. The work has also been integrated in ROS and with University of Manchester's simulated radiation sensor, has been demonstrated in the Gazebo simulator. This line of research opens several possibilities, e.g. the use of autonomous robots for exploration of partially unknown nuclear environments for decommissioning. 29

• Cover