BDI agents are usually implemented using logical and symbolic methods, e.g. AgentSpeak, some robot control is being done by cognitive architectures, like OpenCog. But - if neural networks allows to do the generalization, the reasoning and the language processing - why neural networks are not used for goal-directed (e.g. utilitarian) autonomous agents, e.g. controlling service robots of providing intelligent services more than chatbots can do, e.g. doing automatic programming based on the dialog with the end users?
I asked a bit similar question about continuous learning BDI agents in particular Neural network as (BDI) agent - running in continuous mode (that do inference in parallel with learning)? and received good answer, but I wonder, why Google is not givien recent results for search "neural networks autonomous agents".
I have bad experience with Google, that is why I am asking here. E.g. Google gives very few good answer for general query "grammar extraction from neural network", but Google gives excellent answers for the very specific question "learning context free and context sensitive grammars with neural networks". I feel that there is similar problem with neural networks and autonomous agents - something should be going on, but it uses different, specific keywords, it takes some distinct angle and it is impossible to get good search results because of Google failure to make some semantic search.
This findings http://www.cs.cmu.edu/~dchaplot/papers/arnold_aaai17.pdf and https://github.com/glample/Arnold most likely show that reinforcement learning game playing agents are the current state of the art and there are not much efforts to go beyond them. There are more or less autonomous agents in logical-symbolic worlds that can induce and reason about goal hierarchy (deriving goals from the very general goals, e.g. maximizing utility in some scheme of preferences, earning profit from provided services) and that can act following deduced goals, but apparently - neural networks are not ready for this. But maybe still there are some good research trends?