The Soar Cognitive Architecture: Current and Future Capabilities (John Laird)
In this talk, I review the history of the Soar architecture, including cognitive capabilities such as symbolic hierarchical task planning, execution, and learning. Soar also includes a spatial-visual-system (SVS) that supports modality-specific reasoning beyond what is possible with symbolic representations, which has been critical for robotic applications. Our main challenge problem has been interactive task learning, where a robotic agent learns new tasks in one-shot from guided natural language instruction. We are currently expanding reasoning and learning from events as well as extending SVS so that it integrates with the symbolic long-term memory and learning systems to better support both perception and motor control. The future is about expanding the knowledge in our agents. This includes innate knowledge (core knowledge about space, time, objects, & agents; curated knowledge; NN pre-learned knowledge), experienced-based learning (architectural learning and deliberate learning), and instruction and social learning.