RI Seminar: Leslie Kaelbling : The Role of Rationality in Modern Robotics

Building general-purpose, autonomous robots requires leveraging structural biases—specifically that the physical world is sparse, local, and object-centric. Rather than relying solely on unstructured learning, effective robot intelligence integrates advanced perception with task-motion planning to achieve zero-shot manipulation. By utilizing foundation models as priors for symbolic predicates, robots can infer latent human intent from demonstrations and generalize across new environments. This framework treats robot intelligence as a process of rational decision-making, where agents use causal action models to plan and execute tasks. Furthermore, robots improve through deliberate, autonomous practice, using planning to create scenarios that refine their skill parameters. Addressing challenges like partial observability and memory remains critical for long-horizon tasks, shifting the focus from simple reactive policies to systems capable of reasoning, planning, and continuous self-improvement in complex, real-world environments.