As spotted in this blog post using Erlang in robotics seems a pretty good idea to implement robust systems and concurrent behaviours.

However I do not consider the subsumption architecture (Brooks, 1986) as the best example of action selection architecture for concurrent behaviours. As far as I remember, the subsumption architecture assumes predefined priorities on behaviours (hierarchy), and has a fix set of rules on how some behaviours can subsume others.

As far as reactive architectures are concerned, IMHO Erlang style of concurrency could be more effectively used on architecture using activation networks (Maes, 1989), or a voting mechanism a la DAMN (Rosenblatt, 1995), or a pertinence-based selection like Creature (Blumberg, 1997), or whatever architecture which enable a distribution of control.

But yes, Erlang seems quite a good match to implement autonomous entities (robots or software agents (MAS)).