I Built an Improved AI in Robot Spleef and Challenged the Old Ones to a Team Battle!
ฝัง
- เผยแพร่เมื่อ 13 ก.ย. 2024
- In this episode we build an improved robot spleef vehicle and challenge it against the old vehicles. Will the new and improved AI win or will the old robots still be dominant?
-------------------------------------------------------------------------------------------------------------
About Scrap Mechanic:
store.steampowe...
-------------------------------------------------------------------------------------------------------------
#scrapmechanic #kangaming
What other "personalities" should we add to our robots?
make one fly
What about giving the map some personality? Have some random posts out in the map where the robots will detect them and turn. One of the best parts about having a simple I like seeing how it reacts two different environments.
Edit; to solve your auto turning left when wanting to go straight, you can solve this with a 45° sensor and an XOR gate inserted before you auto left command. (Not gate may be needed)
Sadistic
huh
Rammer and dodger - they prioritize ramming or avoiding other robots regardless of whether there are holes, but will avoid holes when no other robots are present.
If you’re looking to make different personalities, you could have an aggressive one that purposely drives into other robots when it sees them. Or maybe one that looks for one-block gaps and then steers towards them to try and make as many cuts as possible.
He will probably see this comment, so ill leave advice here. Maybe give the bots more sensors? So it can see slightly more. (I mean adding a second set of sensors to the antennas at an angle so it has more peripheral vision)
RAMMING SPEED!
I want big sweeping changes that sometimes work and sometimes fail
He could put color censors on the bots so when a blue sees another blue they turn away from each other as usual, but when a blue sees a red, it rams the red.
@@goop9086 so racist ai bots? 💀
We should NOT have "priority left turning". It should be randomized.
This can be achieved with painting the arena in 2-color noise input that a bottom facing color sensor on the robots can detect. Use that noise to determine if you turn left or right.
Do you have any clue how tedious that would be?
It would also need to be repainted every time, otherwise the games would be almost the same
@@Rocketenginetoaster Not if you do it in big checkerboard chunks. 4X4 being a single chunk. And you only have to color half of it because a color sensor only detects if the block IS or ISN'T the matching color. It's a random coinflip picker where you only have to worry about making the heads side of the coin.
What I don't know is if you can lift-copy-and-paste in challenge mode, which would make it even easier.
And even then, it's the arena, it's not like he's gonna need to make it more than once or is gonna make multiple versions of it. It's a 1-time setup. Unlike the robots that he is constantly tweeking.
Lastly, I think you forget how much kan just doesn't care about tedious painting. Remember his square wheel video where he painted all the ducks individually? And that wasn't even for a sensor, he just did that for esthetics. It was completely optional/unnecessary for the point of the video.
@@admin-ji2pu No, no it wouldn't. The RNG is based on the individual robots POSITIONAL data, which is different every round and every time they need to turn. >_>
Kan made it a point to have each match have a random offset for their starting direction. That alone is enough to offset the random seed.
@@Hadeks_Marow That is actually a pretty good idea, and for the random personalities thing like random turning he could use that mechanic as well so It might be really effective but even though it would be a bit tedious I think you would need to do it as 2x2 at least for randomness sake, especially at the start because even though they have some randomness the start still could be better randomised.
Crazy sensor idea: Back-side robot sensors for strategy:
If you see a robot with your back but not your front, those sensors having a longer reach, turn towards it to cut it off.
As for the hole sensor, maybe have one that's pointed *almost* straight down on both front corners that's linked to the existing ones. That way, you'll catch holes that are right next to the vehicle as well as slightly further out.
All this automation stuff gave me an idea for a multiplayer spleef arena. Add a shootable target to the back of each vehicle. Triggering it inverts the left-right controls for 5-10 seconds.
Randomness can be done using timers. Either click in a pattern using a button, or try to make a 'linear feedback shift register' using a few timers and gates.
Spleef with friends and robots! Big map.
So. . . we aren't gonna do 2 middle sensors so that the robots could "walk a tightrope"? Essentially an inversed line-rider. I feel like that kind of intelligence is important in spleef.
yes this is what i thought aswell
The ramps cause the sensors to trigger (or untrigger rather) causing them to turn their default direction, thus they turn left off the start. Without the ramps they don't seem to be significantly affected by the turning of the platforms. I suspect it's because they hit the pieces that are straight and it causes them to straighten out somewhat.
As for Ideas, my initial ideas for personality would be default right, default left, reduced vehicle detection range, and can't drive straight.
I think a cool personality grouping would be "the avoider" - "the aggro" - "the ignorant" and I don't have a fourth one.
The avoider would work like current blue, the ignorant would fully ignore other bots and the aggro would turn towards other bots as long as there's no hole that way instead of away from them. Who would win that match?
I think the next personality should include a timer that will make it turn left if no sensor for x seconds every y seconds or right if no sensor for x seconds every y seconds to mix things up.
Put sensors in front of the robots to prevent them from driving into narrow holes
you should challenge the guys to give their own bots some personalities
This is a great idea, have them each make their own ai powered spleef bots from scratch (with size and speed limitations)
Not everyone understands logic as well, so they probably wouldn't be able to
@hvtrs8%2F-wuw%2Cymuvu%60e%2Ccmm-cjalngl-UAWlLZ8%7BkALermiSv3eGCXtC I have scrap mechanic on steam but I don’t have my pc set up so it’s aggregating watching what I could be doing. Though I love it
What you could do for the platforms is some sort of sensor that detects when all the robots are off the platform, a wall forms, then it flies away. This should make it so that robots don't go back on the platform
also more severe turns on the platforms for peak randomness
Fantastic long term series idea with the community battles and your own evolutions battles, look forward to it all.
It would be cool to have a random turning left, a random turning right, a random turning both ways and a no random turning in one arena.
What about Sniper spleef? No seats on the car, players are up on walkways above the arena, and have to remote control their spleef drones by shooting buttons on them.
Of course you can also directly shoot the floor, but that means less time managing your drone.
That's actually the most unique idea I've seen in here, so far. I super this idea for Multiplayer Mondays.
@@KalijahAnderson Thanks, I'm glad you like it. 😊
ooh thats a good one
Omg, Kan you are a genius.
I'm of course smarter for pointing out how you just invented Open Plan Maze Spleef.
Take a Spleef map and add random walls with lots of gaps for vehicles to fit through.
Now you can't see what the others are doing whilst also having to plan your route and traps around walls.
This is what happens to me every time i try to improve something. I somehow make it objectively worse.
Especially in scenarios like this one where you are only testing 2 personalities on 4 bots, having 2 left favored and 2 right favored rather than all left favored would help increase variance as bots will react differently to hitting the far edge.
I would love to see a community ai spleef tournament
For random turning, you can just have a loose bearing that has a rod that just barely sticks out, similar to a AD converter. Then with simply just a sensor to detect it, and a timer to choose how often it happens, you have random turning.
And maybe add a sweeping motion to the hole sensor.
My suggestion for the "antennae" with the sensors would be to make them sweep back and forth, stopping when they find a hole so that they have a while to send out that signal that they've found a hole. I did something similar to that when making the center turret on my farm defenses, and I think it might work well here, too.
So we didn't randomize the paint of the arena itself so that we can use a bottom facing color-sensor for a source of RNG? Shame.
We can still do that for a "Chaos Personality".
the robot could just have a 20 hz clock that is sampled whenever it wants a random value
A large spleef arena filled with bots that you compete against alongside other players would be neat but I wonder how well that would work. You could make it single winners, pair people off with a bot, or people vs bots.
Personality idea: "the scared one" if it sees a hole or other robot it just turns away in a 180 turn and runs away from everything!
1:50 welp...that's certainly an interesting, uh, shape.
Yap it's very interesting
I knew at least one person was going to say this
It's nice to see that even with the end of Rooster Teeth, Red vs. Blue still lives on.
Use a timing setup to randomize turning L/R but priorities the sensor. You could also put the hole sensors on a piston with a timer or controller to increase and decrease throughout the game
You could change the starting platforms to all have different settings on rotation so it truly is more random.
Some personality ideas:
Aggressive: tries to cut off others and even ram them
Scared: looks further away and will avoid everything at all costs.
Dizzy: sways left and right instead of going straight
Regretful: will try to make full U turns whenever encountering holes and other bots
And if you want even more personality. You can make some drive and/or turn faster or slower, have different sensor settings, maybe even have some be able to boost, jump, have more guns, or use pistons.
Hi Kan, here ive got an idea to maybe make the robot move a little more natural. how about you spam blocks over the arena of 3 different colors and whenever a bot drivers over suchs blocks they will turn a certain radius.
Yeah I definitely want to see more of this, MM with everyone making a custom sensor/logic setup on the pre-built base would be great
Multiplayer idea: add a turret on top of the robots for the players to control, and maybe give some ability to turn, but add a cooldown to it
A *simple* addition, since you're already using a Controller to set the angle of the sensor antennas, have them loop between angling further down and further up, basically wiggling down and up from where they are currently.
Changing how they interact with other robots might be a good idea. Either make them not care at all about other ones or purposely drive into them.
you should put two sensor antenas facing the front and angled down put them above the end of the wheels so you can make sure that the path isnt too thin) and wire them up so when the side facing antenas both not see ground but both the front antenas are seeing ground it will drive straight and if both the side dont see the ground and one front antena is it will turn to the other direction so it can drive on thin paths (put them above the end of the wheels so you can make sure that the path isnt too thin)
Maybe they need more than 2 directions. They only _really_ need to turn when the path is missing directly in front of them but that's where they struggle most. They dither until it's too late.
Easy suggestion for the next version, make it detect holes directly forward, it would go straight when there are holes on both sides unless the center sensor would detect a hole, then it would go back to default-left
16:30 the robots can’t see in front of them so they drive straight in the hole. So you need a forward facing sensor and if it detects a hole the robots goes left (this could be different in every robot to give them different personalities). The forward facing sensor should of course be lower priority than the side sensors so it doesn’t steer the robot in a hole right next to him but if both side sensors detect a hole and the forward sensor the robot should probably go forward and not turn left. If have a forward facing sensor you can make the side sensors look slightly more to the side (maybe 50°) so they see too the side better. For how far a sensor should look, it should detect the holes just far enough that it can just turn around while turning towards the hole (this happens when all the sensors detect a hole). If the robots looks closer it can’t avoid the hole anymore and it becomes useless to know there’s a hole if it looks farther it maybe looks over a hole it could avoid. This of course means that the side sensors should look farther than the front sensors because of how a turning radius works.
You could have the robots maybe team based where they do different actions depending on what color robot they see? And randomized turning just for variability too would be nice.
I would really like to see these robots get developed more and more :)
Love this! Evolve them to the ultimate spleef machine then bring along to MM spleef! 🤘
Repairing spleef where you can use vacuum pumps to repair gaps in the cars board. Or maybe glowbug spleef? You can drop glowbugs down in addition to the guns?
A lot of your robots falls were from driving directly into a sliver of a hole in front of them where neither left nor right sensors were able to detect it. having a forward central sensor to detect that sort of sliver hole before it gets too close to be able to turn away from them would do a lot to help your bots.
Why is this so fun and riveting to watch?!?
If you want to do real evolution, you need a 1000 parallel instances executing lots of tests so you can get a true statistical winner, not just "I hope best to 3 is enough!"
Maybe a headless scrap mechanic that runs as fast as the engine can process rather than "real time".
For the next evolution have rear sensors to dect other bots. If they get triggered the bot turns toward the other bot. This could lead to them cutting out other bots just because they're behind them
to make it completely fair, you could have four different AI, all spawning on each four sides (so 16 bots), and each side has a different spawn order, in case some variants have slight benefits for being lefter or righter in spawn order. Plus, more AI at a time averages out results better
Instead of rotating plateform you could drop the roots from 20 blocks and let them bounce a little to get the initial randomness
You should double the hole sensors so that when there are holes from both sides the robot will pick this with less" hole" (less double sensors on)
This beginning to become awesome. I hope you continue working on these. Is it possible to add a coloured pinwheel and sensor to the movement checks to add a little randomisation?
12:05 You should add there a repeater or some other mechanism which would determine whether to turn right or left in impasse situation. Ie. when the logic gate is on, turn left, when off, turn right and use some system to flip-flop it. You could make a timer which changes it periodically, or a memory bit so that it always turns the other way than previously, or what ever else. Should be relatively easy to implement.
it is far to exciting see these compete against each other!
I think a good upgrade would be to make it to where if it sees a hole on both sides, but it sees a gap wide enough for its wheel base, it'll drive straight forwards instead of doing the wiggle thing that it does.
The community Sleef Robot fight would be SO much fun to watch
4:11 I think one of the issues the Blue bots were having is at the very beginning of the video you reduced the Hole Sensor angle slightly which may have caused the Blue bots to not see some of the holes until it was too late for them to turn away. I thought I saw some instances during the matches where the Blue bots seemed to get very close to the edge before turning even with no other bots nearby. Might be worth a retest.
Love this idea of a base spleef bot and having others build off it
If you're going to be doing more A:B testing, I'd like to suggest splitting the starts into pairs on opposite sides of the arena. So f.ex team blue starts on the left side, and team red starts on the right. I suspect this would yield much more interesting results, as the robots would start out driving towards each other.
Next upgrade should be having mid front sensor branch or to have the 2 branch pivot less outward which, in this case, I would make the sensor face straight forward on the tire line. The second option is especially good because you actually check if there is a surface for the wheel to go on instead of looking for surrounding holes
Put sensors on all 4 corners facing down to detect holes in conjunction with the other sensors, in combination with
"@Hadeks_Marow : We should NOT have "priority left turning". It should be randomized. This can be achieved with painting the arena in 2-color noise input that a bottom facing color sensor on the robots can detect. Use that noise to determine if you turn left or right."
You could make a Multiplayer spleef where you include an AI robot competing with you guys
Think I might know what could be causing the Large Pistons to flicker:
The spud guns that remove the starting platforms are still firing the entire time. Either make them stop firing after a timer, or attach them to the starting platforms so that they drop as well.
I think you will need another set of forward sensors to look for holes angled between 10 and 20 degrees that are 1 lower priority than the ones you have now to see holes directly in front of them.
If you do set up community events, you could even try making multiple bases for the community to build off off, sort of like battlebot weight classes. Like one is a small electric motor with medium sized wheels, and one is larger wheels with a stronger motor. And even have different sized arenas depending.
Your hole sensors only detect if there's a hole in the diagonal forward directions. There is a giant 45 deg wedge directly in front of the car that can't be seen. You'd need one more viewing directly forward to prevent the bots from driving straight into a hole like blue disappointingly did 2-3 times.
What about trying to make a bot that doesn't mistake thin cuts for dangerous holes? Like maybe give it two hole sensors per side in slightly different angles and they both have to deactivate for it to turn.
You should have one with a sensor in the middle so if it sees a hole in both sides but a path infront it will keep going straight.
It looks like you're almost there... just one more wrinkle to iron out... the 'Hole Vision' acts too much like Goose Eyes and sometimes pays dearly for having too little vision exactly straight in front of it. A second pair of 'Hole Vision' Sensors between the outer ones keeping more eye in the front might alleviate this, though I dunno how to go about Logicing this to co-op with the existing 'Hole Sensors'...
Add a hole sensor in the middle to determine if it should go straight or left.
You might need two to make sure the platform is wide enough to fit and not bottom out
I think the reason why the change doesn't seem very positive, is that it's so small, RNG takes over.
But not all negative things lead to negative things, sometimes. They can lead to even greater things.
I suspect that the ramps cause the bots to turn left at the stat because when they're going down the ramps, it tilts them forward _just_ enough to trigger the 'there's another bot' sensors, which causes them to turn.
Im really excited about this!
If there is any way to do randomness with logic gates (maybe a clock that goes into an and gate), a robot that will “randomly” decide to prioritize one of the sensor could be good
I would add a center set of sensors so when it sees there is no hole and no wall straight forward, it will prioritize going straight and only turn when it sees a hole or wall straight forward
I recon it would be nice to have them randomly decide to turn and such when there is nothing else going on so they don't just drive straight
If you provide a chassis for a community challenge, your login should force that the robot can’t just stop, nor just rotate back and forth in place forever.
You could have like a spinning arm contraption that spins randomly around the arena and spins the bots and then shoots them off so they are in a more random position and direction maybe.
I think a gap ruler instead of binary detection would be good. Check multiple sensors to determine the thickness of a gap. This could allow for decisionmaking contingent on encountering lines, wedges, and the actual edge of the map.
Another neat idea is a robot with timers that alternate between various personalities conditionally.
My last thought is a networked robot that communicates information to its allies.
one without a default turn would definitely be interesting. When in doubt go straight.
Add a personality where when it detects holes or robots on either side it will randomly choose whether to turn left or right instead of auto turning left
Perhaps you could add another pair of hole detecting sensors to the front so that traingle shaped holes don't kill them as easily.
You could put 2 sensors, one facing directly west, the other east, on max range. That way the robot doesn't have to always select left "if all ways are screwed", instead if it doesn't detect any blocks within 15 blocks to the left, but it does to the right, it would turn to the right instead, and vice versa. That way the robot will be more likely to pick the best direction to rotate.
I think having the starting platforms 2 and 2 at 90 degrees all with the ramp things, would be better. That way they would be more likely to meat somewhere in the middle, I think, instead of almost always meeting at the other end of the arena
You could have a fully randomized starting position instead of the platforms.
Simplest, I think, would be to just have some ceiling for everyone to hang of.
To randomize them, add 2-3 beams of different length, connected with a rotating bearing at different speeds and directions, with the robots hanging of a cardboard to be shot off 1block above the map.
Would need to stop the rotation before starting probably, otherwise there might be some throwing of the robots that can cause issues.
i think them being on the same side makes it more intresting, that and maybe have sensors to see if the robots are still on the platform before they drop the start platform may prevent them from getting dropped, and or pistons to pish them off so they dont get stuck maybe? its been years since ive played so idk if the friction or something will hate that but its worth a shot :D
also maybe you can have a seat angled downwards on some pistons in the middle so you can view them easier without the glasses light glare
I'm wondering. What about a 9-Grid arena, and then Everyone gets a player and 2 bots to a team!
"Old Ones" I thought you were talking about eldritch gods there for a sec lmao
I have 2 ideas for how you could make the robots a little better:
- instead of the robot automatically turning left when it doesn't see anything, make it switch to using sensors that are right in front of the front wheels looking almost directly downwards so it can have more precision to be able to avoid holes, and make it default to going straight if there is nothing below the wheels.
- add snail shells to them
"Imma paint all this logic black" leaves one blue
Put the hole sensors on alternating pistons for a bit of rng movement and to give them the ability to see holes a bit more ahead of them.
So, uuhhhh, hey kAN, maybe some slightly longer delays on the rotation for launch pads, cause uh. almost got a swasti there, brother lol
are you gonna do Bots v Players once you have the bots trained up?
You could add some pseudo randomness by sprinkling some colors through the arena (either in a pattern or randomly) and adding a few color sensors to the bottom of the robot to influence the decission making process in certain situations depending on the current readout of the other sensors and the color below. This could make changing the "programming" a bit easier as you would just need to adjust the detected color. You could also color areas of the map in a different color so the behavior of the robot will change depending on the area it is currently in.
Another idea would be a rotating color wheel on each robot and one/multiple color sensors for an on board pseudo random generator. Depending on how slow the wheel rotates you could even hold a state for a few seconds.
Add a support for the center to prevent them from getting cut out
I think they need sensors on the side so when both directions are invalid it check the side to chose the best directions, which should prevent it from going outside the map
They kinda look like snails. You should put little shells on their back😊
I think you screwed over the blues by reducing their hole sensor angle. I don't think they could see the holes as well as the v1 bots
For random turning, you can color some pixels of cardboard, and have the robot turn when it senses them.
What if you were to do a pulse emitter to randomize the starts, or even the direction they pick after the logic contradicts? Then all you gotta do is when they start, flick a few levers, and whatever tick it lands on is the direction they'll rotate
How about have one variation that is like this new one, except it doesn't just avoid holes, but also tries to keep a fixed distance to any hole or cut it detects on either side, moving parallel to any existing cuts; except when it sees a bot on an 180 degrees arc behind it it will switch to cutting thru the the closest line (but not full on hole as that would be suicide) in a wide S pattern to try to drop any bots trailing them?
I think a center sensor would be very beneficial for the logic