Towards a Grounded Theory of Causation for Embodied AI

Taco Cohen

Keywords: causality, intervention, abstraction, agents

TL;DR: We model actions as transformations of a state space, induced by an agent running a policy, and say when such a policy could be called a surgical intervention setting a mechanism.

Abstract: There exist well-developed frameworks for causal modelling, but these require rather a lot of human domain expertise to define causal variables and perform interventions. They are also not grounded in frameworks for autonomous agents such as Markov Decision Processes, nor in classical physics, both of which describe systems in terms of states and time evolution. Existing causal modelling frameworks describe interventions as operations on a model, but give no indication as to which behaviour policies or transformations of state space shall count as interventions. The framework sketched in this paper describes actions as transformations of state space. This makes it possible to describe in a uniform way both transformations induced by a policy and simplified models thereof, and say when the latter is veridical / grounded. We then introduce (causal) variables and define a mechanism as an invariant predictor, and say when an action/transformation can be viewed as a ``surgical intervention'', thus bringing the objective of causal representation & intervention skill learning into clearer focus.