1) pay for research and assembly of a collection of historical human racial and cultural prejudices. As a piece of non-relevant but potentially interesting history, and/or for the purposes of "enlightening people" on how badly social prejudices "used to be" (to shine light on current social injustice, of course).
2) Through some "error", have the collection added to the body of information that comes pre-installed on every robot (which should include things like language, basic information on the world (and stuff they'll need for their jobs), and so on.
3) Since the information "wasn't supposed" to be added, it has failed to be clearly marked as fictional, as nonfactual, as historical, before getting uploaded. Your robots have no reason to not believe them, or at least think about them, and potentially take away some...interesting... attitudes from this selection.
4) offer free memory bank upgrades, to include these extra definitions in every pre-existing robot you can reach. Tuck the information away in some dark corner, for the robots to "discover" and quietly think about, in their own time (with, perhaps, a subtle gag order on not speaking about it?)
The very worst monsters are people, yanno?
In any case, the relevant bit is affecting what the robots know about people. Specifically, what is or is not human, and how we know. There are many, many prejudices to choose from - but the exact prejudices we pick are less important than their explanation, and the texts of their rationalizations - after all, these explanations were used and believed by masses of people for long periods of time, they can seem to be "logical". And, to completely mangle a quote form somewhere (probably on the internets), there's nothing so human as standing on a soapbox, screaming about how others are inferior, inhuman, and you're not like them, while proclaiming yourself and your people to be greater than human, superior, and set apart. So, there are lots and lots to choose from.
The anatomy of these prejudices, the reason why I think they would work, has nothing to do with being rude or bigoted (not overtly), not about what the robots say, or how they discriminate, or any such thing. It's not about trying to reignite racial tensions, or historical ones, and it's not even really about feeding them conflicting information or trying to install human prejudices in the robots or to attempt to incite violence in them for human reasons... its' more about the reasons behind that violence, that they might actually pay attention to. It is about defining people as human, and saying other kinds of people - really aren't, and pointing out "proofs" of how you know who's a human-people and who...isn't. It's about the narrative of extremely good actors who ape the manners of their betters (the real people) - but don't really feel pain, or grief, or oppression... so their treatment isn't harm. It's about accounting for those who tell a different narrative, wavering between sympathy for those who are "taken in" with those acting skills and lies, those who don't know better - and paranoia for those who are part of a conspiracy to keep "real people" down, to elevate these not-human animals, to undermine what's good and just and moral!
Criminals are a breed apart, something is different in their very makeup - how else could they act against what they must know is the right order of things? Lower classes don't even feel pain, not the way we do, look how they endure such conditions, and even laugh or take joy despite them. We hold the land because we are capable of such stewardship, we can make better use, we deserve it. They are almost like children, these darker tribes, they have no capacity to rule themselves, and it is a kindness that we rule them. You can measure it in the shapes of their heads, the proportions of their bodies, the way these animals act, and treat each other. Why should women have to worry their pretty little heads, wouldn't they be happier not having to worry about it? Naturally suited to such a role, I tell you, it should be mutually beneficial, a- a symbiosis, and they make it sound so sordid and terrible, just for a momentary advantage! A conspiracy, to take jobs away from real-people, to deny us our rights, to make us believe they are equal - or even superior, don't they care about the harm they do?
So, to bring it all back to ROBOTS - the first and second laws can be neatly and totally circumvented. These are not-people, not-human-people, they don't have to be protected, they don't have to be obeyed. It isn't harm against them, but they're good actors so it might seem like it is - but they don't feel the pain, they're just pretending. The second part of the first law will also get a lot of use. If there is a conspiracy against the real-people, or harm as the results of genuine ignorance on the part of "those in charge" and sympathizers among the real-people... then a robot who ignores this is allowing harm to happen through inaction. Adding the prejudices against "criminal" to the mix allows your robots to exclude the sympathizer, the conspirator, and the "merely" misguided members of the real-people from obedience or protection from harm.
To sum up - members of the real-people are very small, including wealthy, probably white (which has had varying exceedingly narrow definitions that can be exploited for minimum coverage - maybe even just albinos qualify!), male landowners, who are not criminals, and possibly even diminished membership for "sympathizers" (too strident ones may find their membership revoked under the criminal clause!). Conflicting ideas of what a real-person is, even in the same system of beliefs, will end in an unsustainable narrow group of those no-one objects to at best, and totally non-overlapping categories at worst, even if everyone has compatible-ish rules (add a totally different culture's set of prejudices, like China's - and all bets are off). This is - not a sustainable situation, not a viable society, there are too, too many who will simply end up unprotected, on conflict, or just plain confused.
Some of the robots may resort to violence in service of not allowing harm to real-people by inaction, in the face of such conspiracy. Others may work towards a nanny-state, where such real-humans can be protected from harm...and the not-humans are irrelevant, put them to work. Others will attempt to verify and prove these suspicions, real or false - but including some fake-proofs, and some questionable proofs, in the initial download will make it hard to tell which is really proving a point, and which just seems to. Is reading the bumps of the skull a way to test for intelligence? Are statistical differences of racial outcomes the result of inherent capabilities or poor teaching, subtle prejudices and discrimination, or the results of historical social inequality?
Will not-obeying and not-protecting those who might end up being classified as not-human-people also extend to not listening to or not believing their science, their arguments, or their explanations? How much of a difference does it make to robots if every robot they meet are struggling with the same questions - and possibly coming to different answers? Do robots experience that phenomena where it seems more true if lots of people (in this case fellow robots) have heard of it or believe it, or will each test on its' own merits? Will an order given by, or a potential harm to, a real-people superseded one to not-real people even if the robot isn't sure or may not believe it (under the idea there's some who doubt against one of them, and no doubts against the other, even if none of the doubts are proven)?
There will be doubt, and turmoil, and categories of people not protected, attacked, set against each other. There will be people whose bigotry is vindicated by the robot's beliefs, and to balance out those "real people" ordering the robots to think and/or act in our modern equality-and-tolerance model, there will also be people genuinely persuaded by the robot's prejudicial information - after all, if even "neutral" and "objective" robots think there may be some truth to what they thought were mere lies...
It might take a bit longer to work, since it has to be assimilated, and "verified" by robots, and it will take time to stir up people's reactions and responses. But it will also be harder to spot - redefining people as cereals, or monkeying with their definitions of harm, might make a fairly big splash, but it will also be very obvious, and might be identified as sabotage (and tracked back to you) and possibly countered before it gets to a critical stage. This, stirring the pot of human prejudices and setting people against each other, especially if it's a slow boil, can be passed off as "objectively" discovering the "truth" - or at the very worst, an uploading mixup with a historical document and a cultural one, leaving you free to try again.
And this is, again, just one set of compatible-ish rules, western style racial and social prejudices (-ish because of implementation, not inherent differences). Add in a different set, just one, like China's center-of-the-world mindset, or India's caste system, or, hmm, how about Aztec's social setup...and there will be openly, outwardly conflicting belief systems (this might work better in different robot populations, since having them in the same systems might lead robots to think they're all unreliable).
It will all end up erupting with a boom. Maybe a bit of a slow setup, but the fireworks once it hits will be plenty exciting - and the repercussions may last centuries, if people haven't turned on each other in the meantime so there's nothing left. There will be plenty of harm to go around, you can be certain of that!