HTP Graphics

RAIN PROGRESS // Two phases of advancing benchmarking methodology have been completed, directed at evaluating how meaningful scene representation is, and how robustly state of the art systems perform, all under a single, open-source framework: SLAMBench. SLAMBench can assess the complexity of the representation and how the information is structured - thus, how meaningful the representation is: as an example, representing a scene only in terms of its geometry is less meaningful than representing it as a collection of objects. To understand the robustness of state-of-the- art systems, we performed over 20,000 experiments assessing how they respond to perturbations (moving objects, lighting changes, fast motion) independently, as well as in combination, in both short and long-term operation. Through these experiments, we identified which of the systems can best adapt to new settings, and what distinguishes them from systems which are less adaptable. The development of the new system has so far targeted building meaningful representations and robustness and is able to model the semantics and geometry of both static and moving objects, such as humans. UNIQUENESS // Most methods either focus on a specific metric for a specific task - for instance, the best accuracy of trajectory estimation - or on generality - for example being robust to as many types of perturbations as possible. By contrast, this project seeks to build a system which can be informed by, and adapt, using knowledge about the task. Rather than being robust to all possible perturbations, the system should be able to adaptively respond to the expected perturbations for the environment it is in (for example, indoors or outdoors); rather than always seeking the highest trajectory accuracy, the system may tolerate more or less error depending on context, thus maximizing energy efficiency. FUTURE ASPIRATIONS // The reconstruction system currently requires computational capabilities beyond those of most robotic systems. The development effort is thus targeted towards increasing the efficiency and adaptability of the system, to reduce the computing needs from a desktop machine with a powerful GPU to a low-power robotic platform. Additionally, the accuracy of the system is currently competitive, but not yet highly adaptive. While progress on the adaptability of the system has been made, the final goal is to simply specify the operating parameters for a self-configurable system which will deliver the best performance for the parameters of a task. In the near future, a core aim is to deploy a robot able to navigate and capture the 3D geometry and semantics of complex scenes in a realistic scenario such as autonomously inspecting a power plant or a waste storage unit. Rather than developing new features, the emphasis in this phase is on reducing the computational and power requirements 100-1000 times. This will bring complex 3D reconstruction that currently requires powerful desktops to low-power robots, pushing robot autonomy and environment understanding to new levels. 41 REMOTE INSPECTION

• Cover