Infants do not learn to use their whole body at once. They start with a few controllable pieces, such as eyes, then hands, then reaching, and add more as each becomes reliable. This paper asks whether a robot learns faster when it is taught the same way.
A simulated mobile robot has to find and handle objects in a room, as shown in the figure below. Instead of giving the learner every sensor and motor from the start, a developmental curriculum introduces them gradually. Early stages work with a small, simple set, and later stages add more as earlier skills settle.
The control policies learned at each stage form a sensorimotor state abstraction: a compact picture of the world that the next stage can build on, instead of starting from raw sensor readings. A small set of parameters controls when each sensor and motor is recruited, which lets us compare different curricula against a learner that gets everything at once.
The curriculum made learning up to 10× more sample-efficient on the mobile manipulation task, and the skills learned this way transferred to new domains without retraining.