NVIDIA’s AI Finally Solved Walking In Games — Transcript

Full transcript

1. 0:00You’re probably looking at this and are  thinking: “I’ve seen little 3D people
2. 0:04walking on grids”. Nothing new here,  right? But have you seen this? Oh yes!
3. 0:11So, you are absolutely right, these  walking movements don’t seem anything
4. 0:15new. We have plenty of games already  with moving characters. So what’s new?
5. 0:21Now, you see, in most computer games, these  NPCs are essentially floating capsules.
6. 0:28The game code moves the capsule, and an artist  plays a "walk cycle" animation on top. And,
7. 0:34yup, here’s the problem. If the speed doesn't  match perfectly or anything unexpected happens,
8. 0:42the feet slide on the ground. This makes for  these glorious moonwalking bugs. Or worse.
9. 0:49But, we don’t do that around  here. Not in this paper. Here,
10. 0:53they have these little agents that are not  capsules, but are actually physically simulated.
11. 0:59So, they are not just little recordings being  played over the capsule. That means…whoops! That’s
12. 1:06right. These characters have about 20 motor-driven  joints, and if they don’t control them correctly,
13. 1:14well, this is what happens. Jesus Christ  that must have hurt. Now, continuing…hey!
14. 1:21Oh my goodness, this is even worse. However, look!  Not everything is lost. If you teach them well,
15. 1:28they can still get up. I’ll tell  you how it’s done in a moment.
16. 1:33And it only gets better and funnier from here.
17. 1:36Why? Because with this system, they can learn many  movement types from scratch. No sliding feet or
18. 1:43moonwalking bugs. Although when they go slow, they  look a little constipated. In this case, instead
19. 1:50of going slow, I recommend running like this! Much  better. Hope you make it there in time, little AI.
20. 1:58Now, in most games, to have a crowd avoid  each other, we usually write strict rules
21. 2:03saying that “If neighbor is close,  turn left." Does it work? It does,
22. 2:08but this creates very robotic, stiff,  and predictable crowds. But here,
23. 2:13I feel like this AI-based technique creates  something that looks a bit more organic.
24. 2:19And as a result, it can also deal with a bunch  of different body types as well. Short, tall,
25. 2:26plump, not a problem. I love  how the realism of this system
26. 2:31makes them all move a bit differently. So cool!
27. 2:34Now, because the pathfinding is  generative, their little brains
28. 2:38imagine the route ahead - so they don't just  bounce off walls like confused Roombas. They
29. 2:44actually weave through clutter with a smooth,  predictive flow that feels surprisingly human.
30. 2:50Now, as good computer graphics researchers, let’s  put them to the real test. Terrains! And this is
31. 2:58where the magic happens! Since they are physically  pushing against the world, they naturally adjust
32. 3:04their movement for stairs, and slopes. Even  rocky paths work, and the best part is that
33. 3:10no special "stair animation" is needed here -  this is just pure physics and learning at work.
34. 3:18How the heck does it do all this magic?  Well, Fellow Scholars, meet the brain and
35. 3:23the muscle. Dear Fellow Scholars, this is Two  Minute Papers with Dr. Károly Zsolnai-Fehér.
36. 3:28A technique called Trace is the brain here.  It uses a diffusion model, similarly to the
37. 3:34tech behind those AI image generators.  But here, it doesn’t dream up images,
38. 3:40it dreams up paths. Now this brain doesn’t  just draw a line for a path like in most games,
39. 3:47it imagines. Yup, it imagines the future  and predicts where open space might appear
40. 3:54in the future, and updates its guesses  as the environment changes. Loving this.
41. 4:00Now, we are not nearly done yet. We also  need the muscle. This technique is called
42. 4:06Pacer. Now the muscle doesn’t care at all about  plans. It just looks at the line in front of it,
43. 4:12and is trying to move its joints to stay on  it without eating dirt. So this is why here,
44. 4:19we don’t have a canned animation. Nope. Here,  it is running a live simulation where joints
45. 4:25are fighting gravity every single frame. And  you need to do it 30 times per second. Crazy.
46. 4:33But now here is where the real magic  happens. Here, the brain and the muscle,
47. 4:38they actually talk to each other. If the muscle  is about to slip a rock, it says “holy crap,
48. 4:44help me!” and the brain says “okay,  okay”, and plans a new, safer path.
49. 4:51This constant bickering between the  dreamy brain and the struggling legs
50. 4:55is one of the keys to success here. Hilarious.
51. 4:58But this still doesn’t work. Yet. Don't  think for a second that this "muscle"
52. 5:04came out of the box knowing how to  walk. No sir! The scientists had to
53. 5:09put this AI through a digital boot camp  that would make a marine cry. At first,
54. 5:15this AI is a bit like a toddler on ice. It has no  idea how to use its knees and keeps faceplanting.
55. 5:23To improve this, the researchers use something  called Adversarial Reinforcement Learning.
56. 5:28Essentially, they have a "Discriminator" - a  judge. The AI tries to move, the judge looks at it
57. 5:34and says, "That doesn't look like a human walking,  that looks like a glitching ragdoll. Do it again!"
58. 5:41Now, over billions of attempts, the AI  learns that if it stiffens its leg properly,
59. 5:47it stays upright. And billions is probably  a conservative estimate here. Why? Because
60. 5:53they trained more than 2,000 humanoids  in parallel for 3 days to achieve this.
61. 6:00So it finally also learned that if  it swings its arms just like that,
62. 6:04it looks natural. It is literally evolving the  ability to walk, purely to please the judge.
63. 6:11And it all just happens naturally because the  physics of the world demands it. Beautiful.
64. 6:17And I want to take a moment to appreciate  just how clever the brain really is. We
65. 6:23are used to diffusion models making images  from noise, right? You start with static,
66. 6:28and out comes a cat or something.  Now, here, they start with a noisy,
67. 6:33squiggly line - absolute chaos - and  the model denoises it into a smooth,
68. 6:39logical path. But here is the kicker: just as  you prompt an image generator, they can guide
69. 6:45the diffusion to force agents into social groups,  effectively telling them to walk side-by-side.
70. 6:52Now, I hear you Fellow Scholars asking,  Károly this is cool, but why do we need
71. 6:57this? Games work fine! And I think that is  a fair point. For your average shooter game,
72. 7:02this is not needed at all. But think bigger.  For instance, think about self-driving cars.
73. 7:08When we test autonomous vehicles in simulation, we  need pedestrians. But if those pedestrians follow
74. 7:15simple, perfect rules, it looks too robotic, and  the car learns to drive in a world that doesn't
75. 7:22exist. However, real humans are messy. Some  don’t pay attention to what they are doing,
76. 7:28and some simply don’t care. If a self-driving car  only trains against hardcoded video game scripts,
77. 7:34it’s going to panic the first  time it meets a real human.
78. 7:38Now this system allows us to fill a  virtual city with thousands of agents,
79. 7:44all behaving slightly differently, but all of  them physically grounded. We can tell them to
80. 7:49"walk aggressively," or like a constipated hippo,  and they will generate unique, messy, realistic
81. 7:56data that helps train safer cars. So it’s not  just graphics - it’s a safety simulation too.
82. 8:03And we get all of this for free, not just in  the paper, but the source code is available for
83. 8:09all of us for free. That’s amazing. Subscribe and  hit the bell if you wish to hear more like this.