Task reductions

Open thread to revisit. If task-specific RL policies are like separately built circuits, one per task, what would let one physical skill be translated into another (bike-riding into steak-cutting)? That would be the bridge from narrow policies to a general robot brain.