The two bottlenecks where computer science is needed are sensory kinetics and plan modeling.

Sensory kinetics involve the ability of robot to perceive and react to changing conditions. If you touch a hot stove you instantly draw back, but a robot is not so smart. The latency in the feedback loop which informs the robot of a potential problem and the robot's inability to recognize and react quickly to a threat restricts the applications and performance of robots in many ways. Imagine hitting a pitched baseball. Your whole body is a symphony of motion coming together at the pinpoint of contact. Organizing such complex motions in real-time 3D is currently way beyond the calculational ability of current robotics software. This is a pivotal bottleneck for robotics.

The other key area is quite different: modeling. Whenever a robot is put to work a human needs to set its priorities and show it what to do. How does the human teach the robot what to do? The human must create a model of reality for the robot and show it how to react given the various things that might happen in that reality. Creating software that makes this teaching process simple and straightforward is an open problem. Many common applications that would seem simple to you like sewing are complex enough that they defy attempts at modeling and teaching. The robot might mechanically be able to do something but we cannot give it the knowledge to do so. This is the problem of modeling.