Very interesting talk. I especially liked your example of recursive estimation with priors on a kinematic model. I don't think robotics in particular is stuck, but rather the whole search for artificial general intelligence, mainly because not enough sensor inputs are given. Afaik there is DeepMind research on this, but I think for general intelligence there must be a fusion of many different sensors to enable the AI to build a model of the world that is sophisticated enough for us to consider intelligent. We can create an extremely deep neural network that can classify billions of images, but the fundamental limitation will lie in the input. I think that might be another reason why adversarial attacks against CNNs are quite easy to perform, even black box. If more sensors (input types) are used, these attacks might be more difficult.

With due respect, How is AI-robotics stuck ? Dr. Brock demonstrated that taking time into account is beneficial but thats not a new finding....Spatio-Temporal learning is being used in AI/Machine Learning for a long time.... Features tracking example that he showed is just registration...Also the using K-means to distinguish between two sounds is not even a new thing is machine learning...Multi-Model learning is also being used for quite a few years in AI/Machine Learning...