We introduce Mala, a cognitive architecture intended to bridge the gap between a robot’s sensorimotor and cognitive components. Mala is a multi-entity architecture inspired by Minsky’s Society of Mind and by experience in developing robots that must work in dynamic, unstructured environments. We identify several essential capabilities and characteristics of architectures needed for such robots to develop high levels of autonomy, in particular, modular and asynchronous processing, specialised representations, relational reasoning, mechanisms to translate between representations, and multiple types of integrated learning. We demonstrate an implemented Mala system and evaluate its performance on a challenging autonomous robotic urban search and rescue task. We then discuss the various types of learning required for rich and extended autonomy, and how the structure of the proposed architecture enables each type. We conclude by discussing the relationship of the system to previous work and its potential for developing robots capable of long-term autonomy.