Assuming the robot passes the Turing Test, it can never make it's own decision without extra input. So a robot that can learn based off of it's own past experiences (though the use of sensors and an ability to notice that parts of it's own system failed up doing something) would be the best description of a conscious robot. However, in order for this to be as smart as a human, it would need to go through as as many experiences of every human in the world, and it would take human's too much time to teach it everything we know. A possibility is having robots speak to each other, so that when one learns something, they all do. Such as "don't walk off a cliff in order to continue being intact." At this point, any robot that has received this message will walk up to a cliff, access it's knowledge about walking off cliffs, and then be able to solve the question of "should I walk off the cliff?".
Keeping this in mind, once you have the robot to the point that it can move freely and learn, as well as be knowledgeable enough about it's surroundings that it can make rules based off it, it is conscious. It may still make some terrible mistakes, such as handing someone a knife too quickly and killing them such as in this article, but it will still be conscious. It would be wise to teach it certain things, but it doesn't matter much when we're only dealing with the point in which it becomes conscious.
Because you said that it can change things within itself, that satisfies the need to be self-aware. Being able to change something within itself implies it has capabilities to see it's own code and mechanics, meaning it can identify failures as well.
Likely, and considering at what point we are with robotics, we are waiting for a time that we know the robot will not injure us accidentally (which would set back the robotics industry by many years based on a lack of trust), and combing out the details in ensuring it can learn as necessary, and can be self-sustaining in terms of keeping itself safe. This will likely not be far off.
After all, if it can't perform the three laws of robotics, it shouldn't be performing self-sufficiently among humans.