Spatial Memory and Navigation Ability in a Physically Embodied Cognitive Architecture

Current computational cognitive models of spatial memory only account for few specific cognitive processes, instead of integrating them with a cognitive architecture. Most existing computationally implemented cognitive architectures lack the ability to use spatial information for planning or navigation in real-world environments. Furthermore, there is currently no cognitively plausible model of spatial cognition integrated with an implemented comprehensive computational cognitive architecture that would model spatial cognition and a wide range of other human cognitive mechanisms, while being able to function in the physical world. The aim of this project is the development of a computational cognitive model of spatial memory and navigation based on the LIDA (Learning Intelligent Distribution Agent) cognitive architecture, integrated with the other high- level cognitive processes accounted for by LIDA, and physically embodied on a humanoid PR2 robot with the aid of the CRAM (Cognitive Robot Abstract Machine) control system. The LIDA cognitive architecture will be extended by a conceptual and computational, hierarchical spatial memory model, inspired by the neural basis of spatial cognition in brains. This memory module, representing the environment on multiple hierarchical grids, will be added to LIDA and integrated with other modules such as working memory, attention and action selection in order to facilitate navigation and planning based on spatial maps. The resulting architecture will be physically embodied on a humanoid robot (PR2), to strengthen the cognitive plausibility of the spatial model by comparing its behavior with humans in performing simple spatial and navigational tasks. This will be accomplished by developing an interface between LIDA, implementing high-level cognitive processes, and the low-level CRAM control system, implementing hardware control, visual object recognition, and motor execution. Apart from hypothesis and plausibility verification, this embodied model will also provide a biologically inspired robotic mapping approach that does not need expensive sensors, scales well to large environments due to being hierarchical, and is integrated with important general high-level cognitive functions such as planning, non-routine problem solving and attention. The models ability to navigate in the physical world, and its cognitive plausibility, will be verified in a series of experiments including testing the robots ability to navigate known routes, novel and multi-goal routes; and comparing planning efficiency, planning time, map accuracies, map learning times and other metrics with data from human subjects.

In this project, our goal was to build a model of how brains (and minds) might learn about where things are and how to get to them in other words, how we learn and use a map in our head. Previously, models of how brains learn about space were not able to function in real environments. Robots ignored how the brain works, and brain models ignored the complexities of real environments. We used data from rat brains to look at how they know where they are, despite having bad vision and inaccurate motion senses. It turns out that a model based on statistically optimal localization can explain much of the firing activity in the brain area concerned with location representation. This suggests that neurons in this area are involved in calculating near-optimal location estimates based on inaccurate sensory information. We then built a model of artificial brain cells which might do this kind of calculation, and collected data on how well humans can estimate locations to see whether our model is realistic. Although this simple model matched human performance in small, indoor environments, it broke down when trying to learn large-scale environments such as the buildings in peoples home town. People made much smaller errors than the artificial brain cells. Furthermore, the model could not explain why the biological brain cells sometimes re-activate outside of the location they represent (like daydreaming can replay recent memories, these cells replay recent locations). To explain large-scale error correction, and the observed replay phenomena in the brain, we built an extended model that used the re-activation mechanism to go back and correct any errors based on more recent observations; similar to a taxi driver re-estimating his/her location upon seeing a well-known landmark. We also integrated this extended model with a more general model of how human minds might work, the LIDA model of cognition. We implemented the extended model in a simulated humanoid Atlas robot, rebuilt hundreds of cities in virtual reality, and let the virtual robot learn about them. The robot running the extended model successfully learned about the home towns of over 240 participants, making comparable errors to the human subjects. Apart from learning maps and estimating its location correctly, the model was also able to structure its maps the same way human participants did it knew which buildings belonged together in participants memories of their home towns with approximately 90% accuracy.