The document discusses grounding language and symbols in action and perception for cognitive agents and robots. It describes models where categorical representations emerge from sensorimotor tasks and agents can learn syntactic structures without rules. A multi-agent model simulates population of agents interacting and evolving sensorimotor and cognitive capabilities. Agents in foraging tasks learn food categories through sensorimotor toil or symbolic theft. Robots can acquire compositional categories from basic ones through symbol grounding transfer between agents. Language comprehension in robots integrates vision, speech, and motor processing to respond to instructions.