A Marr-Inspired Framework for Raising “Good” Robots

Our current computer and AI systems are built on Neuroscience principles from almost a century ago. Recent advances in our understanding of biological computation have not crossed into computer science to catalyse advancements. We outline a multidimensional blueprint for a form of bio-inspired agent leveraging modern Neuroscience principles (including the co-localisation of memory and compute, plasticity, embodiment, active inference, and neurodevelopmental principles). We discuss how combining these core features could theoretically lead to cognitive agents that are aligned to our prosocial values, transparent, explainable, and energy efficient (i.e., “good” robots). In particular, we leverage Marr’s tri-level framework and advocate for an “Implementation Level” consisting of embodied neuromorphic hardware, an “Algorithmic Level” consisting of Active Inference, and a “Computational Level” consisting of prosocial goals (supported by evidence of prosociality catalysing the development of our own complex cognitive abilities). A developmental process scaffolds different prosocial computations over time. Supporting our perspective, we include simulation data demonstrating the transfer of priors between two different prosocial behaviours (Computational Level) via Active Inference (Algorithmic Level), supported by an embodied process (Implementation Level). Agent behaviour is transparent and explainable throughout. We advocate for this blueprint as a guide in creating capable, ethical, and sustainable machine intelligence.