The Steve amiibo AI – Research Notes

by Doc – With Significant Help from LilyLambda

I bestow upon LilyLambda +3 Stud Points

I took a metric assload of notes when training the Steve amiibo for the first video training guide. Typically when taking notes on an amiibo, you’ll find one or two interesting facets to take notes on: Steve had eight. I was suffering from information overload by the time I whittled down what I was going to include in the video, so I felt it necessary to produce everything on amiibodoctor.com.

Some of these notes are up to 2 weeks old, so I’ll do my best to explain what I was thinking when I took them.

If you haven’t trained the Steve amiibo yet, you should see the Amiibo Doctor video on him.

Combos

Jab – Jab – Down Tilt – Shield pressure? It certainly looks like Steve amiibo has a built-in shield pressure routine. After taking this note, I saw that there were variations of it as well: sometimes it would only use one jab, or sometimes it would follow its shield pressure routine with a grab if the opponent continued to shield.

Jab – Down Tilt – When not shield pressuring, Steve will often use his Jab as a “get off me” move. We haven’t seen amiibo do this a whole lot since Smash 4, because Smash Ultimate’s amiibo AI hasn’t been fond of jabs for most of its history. It makes perfect sense for him to use a move like Jab that way.

But he’s got six different Jabs – one for each tier of resource, and one without resources at all. So the developers used an easy fix. If Steve uses Jab, and he can connect with Down Tilt, he might try it. This works really well for the lower-tier resources that have less knockback, or when the opponent has little damage. It’s a smart combo option that I wouldn’t have considered.

Up Tilt Chains – I covered this pretty extensively in the video, but I’ll expound here. Steve can finish his Up Tilt chains with Forward and Back air, which is a downright brutal ability. Steve’s aerials are powerful enough without the punishing damage output, so being able to set up a Forward or Back air pickaxe attack while building 30-50% damage along the way is a borderline dystopian combo routine.

Rage messes with Combos – This is broadly true for all amiibo, but it really screws with Steve. Steve’s resource mechanic means that, as I said previously, he’s got six different versions of each attack. Stacking Rage on it means that Steve’s combos are fundamentally unreliable.

Sword combos? N/A atm – When I wrote this, I hadn’t yet come to the conclusion that Steve just didn’t have anything fancy out of sword. In the video, I showed a clip of what I expected – a Steve using Jab repeatedly until his opponent was across the stage and nearly dead. At the moment, that’s not in the AI, but if Steve ever gets patched, it’ll probably be the first thing added.

Later, LilyLambda and I swapped notes and neither of us found anything with his sword (besides the three things I showed in the video), so I’m pretty certain about that conclusion until someone shows me solid evidence otherwise.

Only Does Up Tilt Chains With Wood? – The rumors are true; Doc is sometimes wrong. But I’ll give myself half-credit, because I was partially right.

Steve doesn’t only do Up Tilt chains with wooden axe, he’ll do chains with all the tools. But, because of the way amiibo decide whether they can do a combo or not, Steve does them most often with wood. This is because Wood Up Tilt has the least knockback (besides a raw fist), making it easier for Steve to follow up with the widest variety of percents.

AI Stuff

Perceives edge of blocks as a ledge – This is the weird stuff. I wrote this when I was still training my second Steve at around level 20-30. He had a habit of jumping up high, using a block, and then behaving as if he was edgeguarding me despite being on-stage.

Edgeguarding Shenanigans – On the topic of ledge, Steve’s used a lot of bizarre edgeguards. Sometimes he stays on the ledge and goes for Down Tilt or Down Smash, which is what he did when using blocks in the previous note. Sometimes he goes offstage and tries a Forward or Back air spike, which is pretty impressive.

If Steve is attacked while he’s on the ledge, he’ll do some sort of tech-wall jump-Back air Bullet Time crap. I can’t find a clip of it on my computer at the moment, but you’ll see it pretty frequently if you fight him yourself.

Gets Pineappled on Dream Land & Pokemon Stadium 2 – I cover this more in the Pythra research notes (because my smart ass decided to promise Steve would come first and make the Pythra guide before Steve) but basically the underside of Dream Land and PS2 screws with amiibo AI and prevents recovery, and Steve’s no different.

CHECK REPLAYS – I got nothing.

Doesn’t do TNT on its own – I should couch this statement by saying that I never once used TNT when training my Steve. I don’t see it as a useful move, and it’ll probably end up like a significantly more random and less helpful version of Snake’s grenades if you teach him to use it. Snake’s grenades don’t KO Snake, they’re easily sent to the opponent, and the opponent amiibo AI is hardcoded to get them no matter what.

Steve’s TNT KOs himself at much lower percents than grenades, the TNT block just kinda sits there, and the opponent isn’t really drawn to the TNT.

So I didn’t train him to do TNT, and he never once used it.

Can use mining, but infrequent + Shouldn’t be taught mining – the AI does it best + Makes blocks on its own – because of mining

This is something that I was originally going to put in the video but chickened out and didn’t. I even made a graphic for it:

Steve amiibo mining and crafting explanation.

I chickened out because I’m not more than 50% sure about the theory: but this post just covers research notes, so I might as well explain the theory.

My theory is that Mining, Crafting and Blocks are all governed by Neutral Special’s probability. If you mine more, you’ll place Blocks more. I believe this because amiibo data is formatting according to the move input, and any “extra” moves have historically been handled by the CPU AI. Terry’s Go! moves are a great example of this: he doesn’t have an extra amiibo data slot for the Go! move inputs. Instead, the game hitches the probability on some other move probabilities, and the CPU borrows that probability from those moves. That way you can still train Go! moves by using them more, but it doesn’t break the formatting of the data under the hood.

I figured Steve is the same way. If these three individual moves are all considered the same input, why wouldn’t they have the same input probability?

But there’s a catch: Mining and Crafting are essential parts of Steve, so the devs would probably require the amiibo to at least do those two, and let the block placement be determined the traditional way, but with more favor towards block placement so the amiibo will still use it without significant training.

I know it’s kind of a word salad, and it’s not a theory that I’d put on the channel just yet. I need to test it first by using lots of Mining, no Crafting, and no Blocks to see if it holds. Corner cases are great for testing amiibo AI. If you’ve got some theories on how Mining/Crafting/Blocks interact with each other, shoot me an email at amiibodoctor@gmail.com. Evidence is appreciated.

Path-Block off Ledge + Pillar-Block on Ledge

Even though my Steve wasn’t trained to use Blocks, he sometimes uses them to edgeguard. Here’s a masterfully-produced diagram of where he sometimes places blocks to edgeguard.

Explaining how the Steve amiibo edgeguards opponents

Oddly, when he got to the top of the red pillar, he’d use Down Smash or Down Tilt. I noticed that same behavior in the previous note, “Perceives edges of blocks as a ledge”. I suppose that supports my research.

It’s important to note that he didn’t always go for a full 4-block edgeguard, although that was the most I ever saw him put down.

Omega Stages Changing AI?

Here’s another theory I’m not 100% certain on. I made a short list of things that I believed changed when Steve played on Omega stages instead of more traditional stages.

  • Minecart usage (Featured this in the video because it was the most demonstrable change)
  • Block placement (Steve would never use Blocks at the top of the stage border on tournament-legal stages, but as you see in the video he’d occasionally build a “roof” on Omega stages)
  • Up Tilt combos (this one could be explained by Omega stages just having more airspace to play around with)

There were a few others, but these are the ones I underlined and wasn’t able to provide evidence to the contrary for.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s