Toward an architecture for robots in the era of foundation models
Mohan Sridharan · Edinburgh Research Explorer (University of Edinburgh) · 2025
Robots are increasingly being used to assist humans in different application domains. The ready availability of high-fidelity hardware and data has led to the development of deep networks and foundation models that are now considered to be state of the art for many problems in robotics. However, these methods and models are resource-hungry and opaque, and they are known to pro vide arbitrary decisions in previously unknown situations, whereas practical robot application do mains require transparent, multi-step, multi-level decision-making and ad hoc collaboration under resource constraints and open world uncertainty. This essay argues that to leverage the full poten tial of robots, we need to revisit the fundamental principles that can be traced back to the early pioneers of AI who had a deep understanding of cognition and control in humans. We also need to embed these principles in the architectures we develop for robots, using deep networks as one of manytools that build on this foundation. In addition, this essay briefly illustrates the benefits of this approach by drawing on my work on core problems in robotics such as visual scene understanding and planning, changing-contact manipulation, and ad hoc multiagent collaboration.