The team originally set out to use robots as models for the adaptive, intelligent kinematic problem solving behavior of Goffin’s cockatoos. However, they soon realized that our contribution needs to go beyond “simply” using robots as models in this way, and that they actually need to dedicate significant effort to develop the research methodology itself. They developed and analyzed an approach to identify learning processes given candidate hypotheses. In this study they understood that a common problem in reinforcement learning, in the context of computer science, may also pose a significant problem in analytical studies of learning. To keep it brief, even if they manage to narrow down a list of candidate models, which include the correct model, a naive comparison to experimental data won’t suffice to identify the model. Another challenge that the team identified and started to address is the many-to-one problem. As many competing models could potentially explain our experimental data, a single mechanical model would not be a unique and sufficient explanation of the behavior we see. Rather, they need to make statements about larger sets of mechanical models in relation to the data. This is why they suggest a constraint-based approach where we incrementally gather constraints that may help them narrow down the set of valid models. Some of our research leaned more towards the synthetic side. Inspired by how Goffin’s cockatoos  behavior is robust against unknown and changing dynamics as they manipulate lockboxes, they investigated how we can make robotic behavior robust against unforeseen disturbances and sudden changes in the state of the world. They developed a way to structure behavior such that it can quickly change between discrete behavioral modes and also showed that assumptions about the temporal order of such behaviors is not always necessary and that such assumptions may even be detrimental.