AI/ robotics · machine learning · continual learning · self-awareness

Robots May Develop a Stable 'Self' Through Continual Learning

Robots trained on variable tasks develop a stable subnetwork that acts like a persistent self, and damaging it measurably hurts performance.

Researchers have found that robots undergoing continual learning spontaneously form a stable internal subnetwork - one that behaves, structurally, like a persistent self.

The study compared two setups: one robot trained on a fixed task, another retrained continuously across varied ones. The team defined "self" operationally, as the invariant portion of a cognitive system that changes least through relearning cycles - the logic being that whatever persists through constant change is the closest analog to selfhood. Under continual learning, robots developed a subnetwork significantly more stable than the control (p < 0.001). The pattern held across three different robots, spanning locomotion and manipulation tasks.

The framing is what makes this interesting: instead of asking whether a robot "knows" it exists, the paper offers a measurable proxy - find what the network refuses to overwrite. That subnetwork turned out to be functionally important: preserving it aided adaptation, damaging it hurt performance. The self, if that's what it is, pulls weight.

The paper wisely doesn't claim to have solved self-awareness - but it does give the field a way to measure the question without arguing about consciousness first.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →