RB-Trees are, as well as AVL trees, self-balancing. Both of them provide O(log n) lookup and insertion performance.
The difference is that RB-Trees guarantee O(1) rotations per insert operation. That is what actually costs performance in real implementations. Simplified, RB-Trees gain this advantage from conceptually being 2-3 trees without carrying around the overhead of dynamic node structures. Physically RB-Trees are implemented as binary trees, the red/black-flags simulate 2-3 behavior.
As far as my own understanding goes, AVL trees and RB trees are not very far off in terms of performance. An RB tree is simply a variant of a B-tree and balancing is implemented differently than an AVL tree.