Milo: Mini-Man or Massive Myth?

The gaming press is alive with praise for Peter Molyneux’s latest endeavour. Which is not really a surprise, the industry almost always falls for Molyneux and his affable and enthusiastic oratory. A year later when he delivers the product we usually end up with about 25% of what he promised (which is still usually pretty good) but we all still can’t wait to believe his next pitch. With Milo Molyneaux has claimed that it will be on the cover of science magazines. If it does what he claims it does he is most definitely correct.

What exactly does Molyneux claim he’s got with Milo? He says it is an artificial intelligence that can react and respond to the user’s facial expressions and tone of voice. He claims that “Science Fiction writers, film-makers… haven’t imagined what [Lionhead] are able to do today”. The demonstration seemed to show Milo interacting with the user when she spoke to him. Molyneux told us this was amazing and most of the gaming press bought into it completely. What did we actually see though?

Firstly we saw a polished and very smoothly animated rendition of a boy. Molyneux claimed that Lionhead had been playing with the Natal technology for a few months and yet his studio had managed to put together a demonstration that was more highly-polished and more complex in its animation than the Microsoft first party demos. Milo looked almost retail-ready, albeit in a very short scene.

The first spurious claim is that Milo recognised that his user was nervous by the tone of her voice and her facial expression. This may be true. It is certainly possible (and available in many applications as a method for detecting if a recorded voice is lying) but was there any evidence of it happening in the Milo demonstration? The Milo character reacted with his own rendition of nerves after he was told that his user was nervous and then asked if he was nervous. So it seems more likely that what Milo can recognise is the spoken word “nervous” at which point his animation coding runs its “nervous face” instructions and Milo responds accordingly.

The next thing that Milo showed us was his negligence when it comes to homework. His user asked him if he had finished his school project and Milo looked sheepish. Molyneux told us that this was because the user knew Milo well enough to recognise his sheepishness and press him further on it. How advanced or impressive is this? We have already seen that Milo is capable (as any modern PC is) of recognising voice commands. Was it another case of the animation reacting to a key-word, perhaps “Homework”? What would have happened if the user had chosen to ignore Milo’s sheepishness?

Following on from this the animation turns and walks down a path, unguided and unprompted by the user. She even turns to appear like she is walking beside him. Milo recites a couple of lines of dialogue and makes his way to the end of a small jetty. He throws a pair of goggles to the user who stoops to catch them. The trajectory of the goggles would actually have landed nearer her feet so this is either not a natural reaction to catch the goggles or she’s exceptionally bad at catching. Molyneux tells us this was not acted but I doubt very much that this was the first time his user had been through this set of interactions and animations with Milo.

The user is instructed by Milo on how to put the goggles on. This appears to be an instruction to ensure that she performs an action which can be easily recognised by the cameras. The user then performs a different action (which is pre-emptively prompted by an on-screen icon). The video cuts back to Molyneux at this point so unfortunately we don’t get to see how the goggle-situation unfolds. When it cuts back the viewpoint is moving forward on a scripted path (unprompted by the user) to the edge of the jetty and down, parallel with the surface of the water. The user’s image is projected into the scene and she can be seen interacting with the surface of the water in much the same way as the earliest EyeToy demonstrations.

The presentation is obviously more polished and prettier than the old Sony tech but the reaction of the water seems to be out of time with her gestures. It looks at times like she is following the on-screen prompts rather than the scene responding to her. At one point Molyneux even voices-over that every hand movement is being recognised at just the exact moment that a large upwards gesture is clearly not recognised by the camera.

The next event is billed as “real magic” by Molyneux and has been falsely reported around the internet since it was shown. The user draws a fish with an orange pen on a piece of paper, describing what she draws as she goes. She then passes the paper up towards the camera and Milo appears to take it off her and pull it into his world. He says “orange” so it might be that he can recognise colour (press members who have had hands-on time with Milo say this is the case, he recognises what colour their shirt is).

What Molyneux tells us is that Milo pulls the drawing into his scene after Natal scans it. What we see is the picture held up, Milo pulling the graphical representation of the paper (we can’t see if the image is on the paper) down into his scene and exclaiming “orange”. We do not see if he recognises the drawing as a representation of a fish but (as indicated by Sony’s EyePet demonstrations) that is certainly possible.

That is the finale of the Milo presentation. We were shown absolutely nothing that hasn’t been readily available in computers (and functionally operational on PlayStation 2 and PlayStation 3) for a few years. If you get a chance you should watch the video back again and marvel at how cleverly everything is shown without actually showing anything at all. All of the headlines have come from Molyneux’s voice-over and explanation scenes in the video.

While Milo would be amazing if the claims about him ever came to fruition it is certainly nothing revolutionary or even evolutionary just yet. Even the desire to create him isn’t a new one; Alan Turing wrote a paper posing the question of whether a computer could ever think like a human in 1950 and in the fifty-nine years since then no artificial intelligence has ever passed the Turing Test (fooled a human into thinking they were conversing with another human when in fact it was the AI responding to them). If Molyneux’s claims about Milo ever came to fruition he might be the first computer code to ever pass that test and that feat would certainly get his perfectly rendered face on the cover of science magazines.