I see a lot of stuff arguing we need embodiment.
I don’t think this is the most urgent task. For me the most urgent task is creating machines that model the world, and can query these models for a variety of tasks, integrating a rich apparatus for understanding the world with a rich behavioural repertoire, combining understanding of the environment with skill in an organic way . The recent paper on imagination augmented networks for example was a good use of this approach. Effectively moving out of a behaviorist approach to reinforcement learning, towards something that utilizes sophisticated internal architectures and models the processes and relationships in its environment.
To make things confusing though, some people refer to the account I sketched out above as ’embodiment’, which makes me wonder if the term isn’t losing its meaning.