December 4, 2004

Hard to Believe

by Andrew Stern · , 3:23 am

Robin Hunicke attended last week’s Game Tech industry seminar, I assume circumventing the $2450 registration fee :-). The gathering was comprised of a Creating Believable Characters Seminar and a Game Tech Leadership Summit. She wrote up a great three part summary of the event. (Update: Make that five!)

Robin reports that the believable characters seminar was pretty much limited to (impressive) animation techniques; the presentations went little into AI and behavior, because there’s little tangible work to talk about there.

…about AI and believablity, it’s clear that they tried to find a good speaker or two – and just couldn�t. It�s not that people aren�t trying some simple things… or even that they aren�t attending the conference. For example – Checker (at Maxis) and Jay (at Valve) had a long debate during a break on Day Three about whether the industry is ‘doomed’ because for all our realism, characters are still empty husks. So clearly, it�s being discussed. But results are limited, work is slow, and not a lot of people are stepping up to say what they think will take us in the right direction. That worries me.

Um, hello… (pdf GDC04 powerpoint GDC04 video)

But, okay, generally speaking, Robin is right — no group has yet built a working demonstration, let alone entertainment experience, with a broadly capable, non-shallow believable interactive animated character.

By broad I mean characters who have a full repertoire of the basic actions and reactions you’d expect a good actor to have, from walking and sitting and running and using objects, to eating and sleeping and playing, to emoting and gesturing and talking and acting dramatically. And, doing all of those things with personality, varying emotion, charisma and character. It’s not enough to have an animation loop or two for each of these; we’re talking hundreds of reasonably complex, intermixing behaviors and procedural animations. (And to keep things simpler I’m not including natural language understanding or generation here, or fancy ragdoll physics.)

With the ABL language technology of ours I just linked to, we’ve built slightly broad, non-shallow conversational characters (but ones not yet publicly released — we’re almost done!), but even so, they only have limited, shallow physical action behaviors. The good news is, as our paper and talk suggest, ABL is very much capable of supporting rich physical action, but, we haven’t done the hard work to build those behaviors yet. And it’s a lot of hard work, even with ABL in hand. (And it can eventually be in your hands, once ABL gets publicly released.)

Bryan Loyall, who created the original Hap believable agent language (the language ABL is based on) as part of the CMU Oz Project (scroll down to his thesis) has continued his work at Zoesis. Zoesis has released a few believable agent demos over the years, but frankly they don’t show off the capability of their technology well enough yet.

Some games have some impressive custom character behaviors — games like Ico, Half Life 2, some of the better sports games, etc. come to mind; but they are limited to just the few behaviors the game needs; these characters aren’t broadly capable.

Petz (I was one of the developers) are somewhat broadly capable, not-too-shallow dog and cat characters, but they are only dogs and cats, and they aren’t really fleshed-out detailed characters. The Sims are surely the most broadly capable interactive characters built to date, but their behavior is pretty shallow, sometimes ant-farm like.

There’s a bunch of research labs working on the technology needed for deeper believable characters, and some under-the-radar companies and groups working on various pieces. But we’re still waiting for a fully-realized demo to emerge from all this work.

A thought experiment: what is it really going to take to create broad, non-shallow believable interactive characters?

If someone said to me, Andrew, tell me what you need to put together a team to do this — a Manhattan project for believable characters — here is what I’d say I’d need, to create just one really, really good believable character:

This would probably cost about 3 million dollars.

Once the first character has been created, it would probably only take 25% of the original effort to make an additional character. And so on.

This would be difficult, relatively risky R&D.

Add in basic natural language understanding and generation, and we’re talking 24-36 months of 3-4 talented, experienced engineers and writers, starting with the best of today’s NL technology, adding about another two mil to the budget. (The $5 million virtual man… we can build him… we have the technology…)

Well… I guess Robin has reason to be worried.

(hey, are there any wealthy frustrated gamer philanthropists out there reading this? :-)