Architecture integration of System 1 and System 2 in embodied agents

Design and characterize an embodied agent architecture that integrates fast, intuitive System 1 policies with slow, deliberate System 2 reasoning, determining whether components operate in parallel or via on-demand handover, when each response is appropriate, and whether the overall architecture is two separate systems or a continuum.

Background

Drawing on Dual Process Theory, the authors propose that robust, versatile, and safe real-world robot performance will require combining intuitive, efficient learned policies (System 1) with deliberate, structured reasoning (System 2).

They emphasize uncertainty about how these systems should interoperate in practice—whether concurrently or with explicit handover—and whether they are fundamentally distinct modules or points on a continuum. Determining this is crucial for building safe, adaptable embodied intelligence.

References

How to design an architecture that effectively combines the best of both worlds remains an open and potentially fruitful research question~\citep[e.g.,]{goyal2020inductive}. It is not clear whether components of both systems run in parallel or run on demand with an explicit handover between deliberate planning and low-level intuitive policies. In analogy with metacognitive challenges faced by humans, when a System 1 response is appropriate over a more deliberate deployment of System 2 remains one of the open questions to be addressed. Furthermore, it is not clear if the architecture is two separate systems or is in fact a continuum or tight integration of processes capable of fulfilling either part.

— From Machine Learning to Robotics: Challenges and Opportunities for Embodied Intelligence  (2110.15245 - Roy et al., 2021) in Section 3.1 (Robots Thinking Fast and Slow: Limitations and Challenges)