Show HN: Quake 1 movement physics reinforcement learning project(github.com)
github.com
Show HN: Quake 1 movement physics reinforcement learning project
https://github.com/matthewearl/q1physrl
20 comments
I don't know what I just watched, but these movements seemed not very physically accurate to me.
The goal is to push the limits of the game's physics not real life. Quake based games allow you to retain speed when jumping at the same 'frame' as landing (bunny hopping) and slide up slopes with enough speed.
These physics create a high skill ceiling for speed runs and in normal play because of how powerful movement becomes during play.
These physics create a high skill ceiling for speed runs and in normal play because of how powerful movement becomes during play.
Karl Jobst has a Youtube video covering the speedrunnning record history of Quake's first level, which serves as a good overview of Quake movement technique:
https://www.youtube.com/watch?v=43d8fICz6gM
It starts with an explanation of the significance of Quake to speedrunning as a whole, so skip to 5:46 if you only want to see the movement tech.
https://www.youtube.com/watch?v=43d8fICz6gM
It starts with an explanation of the significance of Quake to speedrunning as a whole, so skip to 5:46 if you only want to see the movement tech.
This is an incredible video - thanks for sharing. Couple notes for those like me who are new to the speed running world:
* The goal here is to get through a particular level as fast as possible. Not to kill enemies or get points; it's just speed.
* The physics engine computes a speed for your player. There are some movements that increase it and some that decrease it. What these speedrunners are doing are (1) finding techniques in movement (2) exploiting features of the map that result in increasing their speed and maintaining it.
* These techniques include running along a wall (unclear to me why this increases or maintains speed), bunny hopping (jumping and turning at the same time in such a way where you don't pick up friction), leveraging features in a map to give you added speed (blowing up an explosive barrel and using its shock to speed you up).
* Over the course of 20 years, players have managed to get the record from 30 seconds down to its current 20 seconds for the quake map in question.
As someone who spent many hours playing Quake 2 on multiplayer, the best players moved in a rhythm similar to how the speed runner players do here.
* The goal here is to get through a particular level as fast as possible. Not to kill enemies or get points; it's just speed.
* The physics engine computes a speed for your player. There are some movements that increase it and some that decrease it. What these speedrunners are doing are (1) finding techniques in movement (2) exploiting features of the map that result in increasing their speed and maintaining it.
* These techniques include running along a wall (unclear to me why this increases or maintains speed), bunny hopping (jumping and turning at the same time in such a way where you don't pick up friction), leveraging features in a map to give you added speed (blowing up an explosive barrel and using its shock to speed you up).
* Over the course of 20 years, players have managed to get the record from 30 seconds down to its current 20 seconds for the quake map in question.
As someone who spent many hours playing Quake 2 on multiplayer, the best players moved in a rhythm similar to how the speed runner players do here.
Speculating here based on my experiences in the past building my own physics engines: running along a wall could speed you up a little bit depending on how they’re handling “keeping you outside the wall”. One common/simple way is to apply x1 = x0 + v*dt, calculate collisions after the moment calculation, and then fudge your position back to being inbounds. If I recall, Q1 used simple bounding boxes on all of the entities as well and there’d probably be some (pardon the pun) corner cases in the “keep out of the wall” algorithm.
This is great, I always wondered if one could train an AI to play god level quake as they've done with DOTA and other games since it's such a quirky engine with so many niche and semi obscure moves and tactics.
For anybody wishing to try or revisit the game you can download the nquake pack [0] which a free and ready to go version of the game that includes the shareware portion of the original game and some gpl'd maps that are the most commonly used (if you own the original game you can just copy the \id1\pak1.pak file to the install folder and delete the gpl maps pak). Also included is the ezQuake [1] build which is a modern implementation of the engine that everybody uses these days and features lots of tweaks for improved graphics, fps, teamplay, better menus, better server browser and a lot more.
Then you can use a website like badplace [2] where you can look for servers and see who's playing. From there you can join or just watch some games by clicking the observe or QTV links (if you registered the qw:// handler during install/first launch).
It's still a brilliant game and there's still a relatively decent community playing everyday although the skill level is of some of these players is truly insane.
[0] https://www.nquake.com/
[1] https://ezquake.github.io/
[2] https://www.badplace.eu/
For anybody wishing to try or revisit the game you can download the nquake pack [0] which a free and ready to go version of the game that includes the shareware portion of the original game and some gpl'd maps that are the most commonly used (if you own the original game you can just copy the \id1\pak1.pak file to the install folder and delete the gpl maps pak). Also included is the ezQuake [1] build which is a modern implementation of the engine that everybody uses these days and features lots of tweaks for improved graphics, fps, teamplay, better menus, better server browser and a lot more.
Then you can use a website like badplace [2] where you can look for servers and see who's playing. From there you can join or just watch some games by clicking the observe or QTV links (if you registered the qw:// handler during install/first launch).
It's still a brilliant game and there's still a relatively decent community playing everyday although the skill level is of some of these players is truly insane.
[0] https://www.nquake.com/
[1] https://ezquake.github.io/
[2] https://www.badplace.eu/
Wild, I was just playing QW a week ago for the first time in 20 years and someone was showing me how to do this running technique after destroying me in a one on one.
This is awesome -- I grew up playing Quake 1 and really fell in love with it more as we as a community learned to exploit the movement physics. Still my favourite FPS and still arguably the one with the steepest learning curve.
I didn't have a chance to watch the video yet, but out of curiosity does this actually use QuakeWorld or base Quake 1? As I remember the physics are different between the two, right?
I didn't have a chance to watch the video yet, but out of curiosity does this actually use QuakeWorld or base Quake 1? As I remember the physics are different between the two, right?
It's the Quake 1 physics, not QW. I haven't looked at the QW physics in detail, but I think the main difference is that there's a single frame of ground friction applied between every jump, which ends up making a difference to the optimal techniques.
It's difficult to tell, but my guess would be quakeworld.
QW has slightly faster strafing, so there is a significant difference.
QW has slightly faster strafing, so there is a significant difference.
Given the fully trained RL model, would it be possible to infer and explain the optimal movement technique in terms of a simple rule set?
How difficult would it be to encode the optimal running technique as a rule instead of learning it with RL?
It wouldn't be difficult at all if the optimal running technique was known before hand; I think the goal of many of these RL exercises is to either i) find a better solution than what we may have imagined or ii) confirm that our knowledge was indeed the best possible solution!
> It wouldn't be difficult at all if the optimal running technique was known before hand;
You're right, but this is a simulated environment. It means the physics in it are driven by a small set of known (and most likely deterministic) rules.
So I wonder if it would it be possible to analytically determine what is the best possible running strategy.
You're right, but this is a simulated environment. It means the physics in it are driven by a small set of known (and most likely deterministic) rules.
So I wonder if it would it be possible to analytically determine what is the best possible running strategy.
Correct, it wouldn't be hard to take the best known strategy and implement it perfectly. Aside from the two points you mention, it also serves as a precursor to applying RL to real levels in which we may not be so sure about the optimal strategy.
I just thought that's the geekiest thing I've seen in a while... good job! :)
That's a really cool project! And it made me want to play Quake again.
http://fte.triptohell.info/demo
There ya go!
There ya go!
Now do CS:GO!
Human run [1] vs scripted bot run [2].
This is from DFWC, a race series that takes place every second year, and using modified Q3 engine.
There are 5-7 maps per season, and people race from start to finish, and best time wins.
Also, each round there is scripted run, to show how far you could push the engine physics, if you had inhuman reflexes. Highly recommend browsing entire YT channel, it's really awesome, if you ever played Q3!
[1] https://youtu.be/FUguBKICPVk
[2] https://youtu.be/g28ObPwPE34