Is Storage a Real Issue for CL in Robotics?

It is true that storage is not a problem now, but think about storing not only small images, but multi sensory data, such as video frames + audio + radar + lidar + …
Memory, especially in embedded devices, will be saturated quickly.

Moreover, even if we memorize only images, an agent that is able to continually learn for its entire life sooner or later will reach a point of saturation, no matter how big the memory is. Just think about a robot with a camera in a room, which have to learn to recognize all the objects that it sees, and more objects are continually added or removed to the room. If the robot only relies on memorizing sensory data and replaying them to avoid catastrophic forgetting, after some time it will certainly forget some object that has not been shown to the robot for a long amount of time.

It’s true that nowadays the major problem with replay is computation, but if we want to build continual learning agents that really learns during their entire lifetime the memory required could be an issue.

1 Like