The Low-Tech AI of Elden Ring

FromSoftware's Elden Ring uses a low-tech AI system based on a pushdown automaton implemented in Havok Script, where NPC behavior is driven by a stack of Goals that can push sub-goals and unwind on failure. The AI relies on weighted random selection of actions during Goal activation, offering a surprisingly simple yet effective approach for complex enemy encounters.

The Low-Tech AI Of Elden Ring FROMSOFT has a reputation for diverse and punishing npc encounters across the entire Soulsborne extended series, but the implementation of the AI decision making itself is perhaps unexpectedly low-tech. Since the majority of the code is implemented in Havok Script A games-oriented Lua implementation from Havok it’s pretty easy to take a peek behind the fog wall to see how they’re implemented. Note that none of what follows is original research, I’m just reading the code that others have done the hard work of extracting, decompiling, and reversing. Goals The primary tool of the FROMSOFT AI approach is the Goal 1 fn1 , which is their own terminology for a unique state that the AI can be in. Goals can be parametized when instanciated, and can access data stored on the Actor itself, but are otherwise really just an immutable table of functions. Now the simplest option would be to organize states into a Finite State Machine https://en.wikipedia.org/wiki/Finite-state machine or maybe a Hierarchical Finite State Machine, but FROMSOFT go one step further and give the system a stack of states. This turns it from an FSM https://en.wikipedia.org/wiki/Finite-state machine into Pushdown Automaton PDA https://en.wikipedia.org/wiki/Pushdown automaton . That’s an entirely abstract definition, so after you get back from wikipedia let’s talk about it concretely from the top down. Each frame Actors will update the Goal on top of their stack of Goals. When the Goal updates, it can then push more Goals as Sub-Goals onto the stack, the topmost of which will execute next frame. The Goal’s update function returns a value indicating either Continue, Success, or Failure. Continue will leave the stack unchanged, the other two will cause the Goal to be popped from the stack. Failure will additionally cause all other unexecuted Goals to be popped from the stack up to the parent Goal The Goal which pushed this sub-goal . For example, we might define a Goal called CoolBossBattle , during the course of its execution it might then push a series of Attack Sub-Goals. Those attack Goals can be parametized by various means, but the main one is the animation id 2 fn2 . GOAL STACK 3: Attack R2, Combo <<<<-- Currently Updating 2: Attack R2, Repeat 1: Attack R2, Finisher 0: CoolBossBattle After a few seconds the first attack lands, and that Goal completes with success and is popped from the stack. However the next fails, causing the stack to unwind to its parent. GOAL STACK 2: Attack R2, Repeat <<<<-- Failed, will be popped from the stack. 1: Attack R2, Finisher <<<<-- Will be removed as well. 0: CoolBossBattle Readying it to chose its next action now that the attempted combo of attacks has ended. GOAL STACK 2: Attack L1 1: Attack L1 0: CoolBossBattle <<<<-- Updating, pushes 1, and 2 for the next frame. Not too complex 3 fn3 In their APIs they refer to the root of this stack as the “Top Level Goal”, which I’ve made confusing by referring to the currently executing goal as the “top” of the stack. So keep in mind those are separate things. Activate Goals are defined by a few functions used as callbacks, and the one which contains the most AI logic is usually activate. This is called the first time that a Goal is updated, and then every subsequent time that the Goal exhausts its Sub-Goals and starts executing again. For boss and regular npc Goals the code in Activate is responsible for choosing the next action that the Actor will take using a mix of context from the world and Actor, and randomness which also comes from the Actor itself . The most widely used approach uses common code to perform a weighted random selection between a number of Actions which are just functions , calling the winner. To return to our CoolBossBattle , this time in some Rusty pseudocode… fn action giga death ray goals: &Goals, actor: &Actor { todo ; } fn action leap attack goals: &Goals, actor: &Actor { todo ; } fn action ground slam goals: &Goals, actor: &Actor { todo ; } fn action light attack combo goals: &Goals, actor: &Actor { let target distance = actor.target distance Target::Enemy ; let fate = actor.next random ; // ApproachTarget itself being a goal defined in common code if target distance 2.0 { goals.push sub goal Goal::ApproachTarget, Target::Enemy ; } goals.push sub goal Goal::Attack, AnimId::R1, Combo::Initial ; goals.push sub goal Goal::Attack, AnimId::R1, Combo::Repeat ; // Unlucky buster It's the long combo. if fate < 0.2 { goals.push sub goal Goal::Attack, AnimId::R1, Combo::Repeat ; } goals.push sub goal Goal::Attack, AnimId::R1, Combo::Finisher ; } fn action heavy attack combo goals: &Goals, actor: &Actor { todo ; } fn activate &self, goals: &Goals, actor: &Actor { let target distance = actor.target distance Target::Enemy ; let mut weights = if target distance 6.0 { 15.0, 65.0, 0.0, 10.0, 10.0, } else if target distance 1.5 { 0.0, 0.0, 5.0, 60.0, 35.0, } else { 0.0, 0.0, 20.0, 40.0, 40.0, }; // This doesn't exactly work this way in the Lua code, and these cooldowns // don't make sense either, but hopefully it gives the rough idea. // // The helper function is checking last played data for the animation on the // Actor itself, and then modifying the weights before they go into the // common battle randomized selection. weights 3 = if common::is cooldown goals, actor, AnimId::R1, 8.0 { 0.0 } else { weights 3 ; }; weights 4 = if common::is cooldown goals, actor, AnimId::R2, 10.0 { 0.0 } else { weights 4 ; }; let actions = action giga death ray, action leap attack, action ground slam, action light attack combo, action heavy attack combo, ; // Does some common setup for the number of actions and then rolls the dice // and chooses which function to call. common::battle activate goals, actor, weights, actions ; } Modifying the weights dynamically is handled in many different ways, but the most common are simple rng rolls from the actor and hp thresholding. Other, simpler, Goals than the top level battle Goal for an Actor may simply push a few sub-goals, perhaps reading some data from the Goal parameters. The nesting means that it’s possible to compose quite complex behavior from simple building blocks. Interrupts The other major callback defined for goals is the Interrupt. As the name suggests, this allows Goals to respond immediately to external events which are mostly configured on the Actor itself. My understanding is that interrupts bubble up, that is, it will run the interrupt on the currently executing Goal and then its parents recursively, until it runs out of Goals or one of the interrupt callbacks returns true to indicate it has consumed the interrupt. For example, if I wanted CoolBoss to move into a furious rage of attacks as soon as I set it on fire, then I might implement something like the following. fn interrupt &self, goals: &Goals, actor: &Actor, interrupt: Interrupt { match interrupt { // If I start burning, attack SpecialEffectActivate { target, special effect, } = { if target == Target::Self && special effect == SpecialEffect::Fire { // Since there might still be other things running when // interrupt is called we need to unwind so we're on top again. goals.clear sub goals ; goals.push sub goal Goal::Attack, AnimId::R1 ; goals.push sub goal Goal::Attack, AnimId::R2 ; goals.push sub goal Goal::Attack, AnimId::R1 ; goals.push sub goal Goal::Attack, AnimId::R2 ; return true; } } // If somebody uses an item they might be in for it. UseItem = { let fate = actor.next random ; if fate < 0.5 { goals.clear sub goals ; action light attack combo goals, actor ; } } // Perform a ground slam if I get attacked from underneath. Damage { target, } = { if target == Target::Self { let distance = actor.target distance Target::Enemy ; let fate = actor.next random ; if distance < 1.0 && fate < 0.8 { goals.clear sub goals ; action ground slam goals, actor ; } } } = {} } false } This is used to implement some truly evil features, for example the Bell Bearing Hunter will detect you spell casting or using an item and from there has an 85% chance to immediately abort its current action and launch into an attack. They also make use of dynamic spatial watch regions configured on Actors, which trigger interrupts. For example you might add a watch for the area behind or under a boss, and use that to adapt their behavior immediately when the player tries to get clever. Timeouts Goals, in addition to their individual state, carry a lifetime value in seconds. This is used to break out of states which become stuck for whatever reason, and lifetime seems to be used mostly as a bug containment mechanism. It’s also possible to modify the lifetime of a parent goal during execution, to indicate continued forward progress. Actor Data Access In many AI decision systems you might have heard of fancy systems for data storage like “blackboards”. In the Souls games there’s an array of floats on each Actor which are set and read arbitraily from Goals by index. Good enough I suppose A callback I didn’t mention before, Initialise, is commonly used to reset this data when an Actor is assigned a new Top Level Goal. Goals have access to a range of queries about the world through the Actor. As far as I can tell most of these are pretty “low cost” from a performance perspective. Aggro and Targeting seems to be handled outside, so it should be possible to keep the Goals very lean even considering it’s all interpreted Lua. Actual Doing Stuff Something I’ve entirely skipped over is how the Goals actually Do things. For the most part everything in FROMSOFT games is animation driven. The Goal says “play this attack animation”, and then the animation events carry hitbox information and timings, special effect triggers, projectile creation events, and whatnot. They also have a variety of “combo” features which seem to boil down to choosing a different set of events in the animations to enable faster linking of chained animation during a combo attack. At some point they went all-in on Havok middleware. The animations are authored with Havok Animation Studio discontinued . Previously we mentioned the AI scripts are using Havok Script also discontinued . Physics is handled by Havok’s physics, and pathfinding is delegated to Havok AI not discontinued, but renamed to Havok Navigation . Misc Stuff They seem to split AI scripting into a “logic” script, and a “battle” script, where the logic script is far more sharable, and the battle scripts are often bespoke. This seems super smart, it’s common to run into issues jamming both these things into singular hierarchies. Level designers are able to configure the Top Level Goal for an Actor in the level itself, so you can place some enemies down with a passive Goal rather than their usual combat Goal, and they would just chill whilst otherwise functioning normally. Most of the common code is relatively compact bits of Lua, but I believe load bearing Goals like Attack and MoveToSomewhere are implemented in C++ which gives you a pretty nice balance of scriptability and performance sanity.The update function itself is sometimes used to check conditions, I expect this must have caused problems occasionally. But so long as the interface for Actors in scripting is thin I guess you can keep it under control. Don’t add a pathfind function call… I’ve entirely skipped over the event scripting system used to do high level encounter logic and level scripting. Unlike the AI it seems to be entirely custom, with a very restricted VM. That said, since it’s not Lua it’s hard to see how they’re actually authored. If anyone knows of primary sources for info about their tooling that would be super cool Conclusion There’s a lot of enduring hype for complicated AI systems GOAP https://www.gamedevs.org/uploads/three-states-plan-ai-of-fear.pdf springs to mind but I think the success of putting a lot of control in the hands of your designers and animators really speaks for itself. A pushdown automaton https://en.wikipedia.org/wiki/Pushdown automaton is also fundamentally quite fast compared to Behavior Trees https://www.gameaipro.com/GameAIPro/GameAIPro Chapter06 The Behavior Tree Starter Kit.pdf and planners. Behavior Trees https://www.gameaipro.com/GameAIPro/GameAIPro Chapter06 The Behavior Tree Starter Kit.pdf often require top-down re-evaluation of a complex tree of scripted nodes, whereas this is almost always executing a single Goal from the top of the stack 4 fn4 . Planners like STRIPS https://ai.stanford.edu/~nilsson/OnlinePubs-Nils/PublishedPapers/strips.pdf , GOAP https://www.gamedevs.org/uploads/three-states-plan-ai-of-fear.pdf and HTN https://en.wikipedia.org/wiki/Hierarchical task network add an expensive search to the middle of everything. Compared to FSMs https://en.wikipedia.org/wiki/Finite-state machine the flexibility of dynamic transitions makes it easier to avoid an explosion in the number of states and their transitions. This also makes it far more reasonable to compose AI functionality in an imperative way. Plus of course it’s dramatically more legible than planner based solutions where individual actions are moved out of the hands of combat designers. Is it going to handle more complex scenarios than the typical Soulsborne npc or boss fight? I actually think it can go quite far. Update Some hackernews commenters were confused by the comparison with BTs, pointing out that this scheme is quite similar to an event-based BT which keeps a stack of active nodes, and wondering about the complexity compared to “normal” game AI. Since I kind of phoned it in explaining all that stuff I’ll try to expand a little. Firstly while it’s true that event-based BTs do avoid the requirement to perform a top-down re-execution of the entire tree, not all BT implemetations actually work this way Especially the more academic and early users look a more like the naive implementation. I find the approach here which explicitly represents the execution structure and builds it in imperative code to be significantly more straightforward than trying to retrofit a kind of execution path cache on top of a BT. Secondly, with respect to the broader question of complexity, BTs implement control flow structures inside the BT structure as nodes again, especially so in academic implementations , this is why you see BT terminology like “Conditions”, “Sequences”, “Selectors” and “Parallels”. This significantly bloats the size of trees, and the complexity of authoring them in my limited I’m not an AI programmer experience. By comparison, the FROMSOFT style, even today, has an extremely low state count, and relies on imperative code inside those states to implement the majority of control flow. This, with my performance hat on, is tremendously important for avoiding death-by-a-thousand-cuts style issues where too much logic is held up in gnarly tree structures that tend to be hard to manage during authoring and execution. And finally, for large scale games the amount of code here is just low. FROMSOFT rely on a relatively large number of bespoke behaviors one or more for each boss, practically , but those behaviors are quite small compared to basically anything you’d expect from other large AAA games. In production BTs elsewhere one wouldn’t be surprised seeing trees of tens of thousands of nodes, on top of hundreds of individual nodes implementing control flow and actions. Similarly aspects like the data model on Actors is just extremely barebones here compared to the rich data you’d perhaps expect to see in a “modern” BT implementation. To reiterate a point from earlier as well, planners are just complex compared to this, and FSMs often end up with a similar node explosion to BTs. Where I talk about performance it’s important to consider it in the context of implementation requirements. A “weekend” long implementation of the structure used here on top of a generic scripting language is going to be basically sufficient to implement a game like Elden Ring. 5 fn5 But if you want to build a behavior tree implementation that hits similar performance goals, you’re talking about building a system that lowers the tree as authored to an optimized representation of some kind, on top of implementing a whole host of basic control flow and data passing primitives, plus all the tooling required to author, compose, and debug trees. The naive implementation is likely to be insufficient. So can you achieve good performance basically however you choose to design a system like this? Yes. But as I see it you’re going to need to do a lot more work going the BT route. References Most of the info in this post comes from eladidu readable ds lua https://eladidu.github.io/readable-ds-lua/ it’s fantastic and you can find many interesting definitions as well as a little tutorial. If you want to get even more excited there’s a bunch of tools for extracting data from the game packages, as well as nice modding tools for patching things here and there. This is not to be confused with the concept of a goal which you might know from advanced planning systems like STRIPS https://ai.stanford.edu/~nilsson/OnlinePubs-Nils/PublishedPapers/strips.pdf , GOAP Goal Oriented Action Planning https://www.gamedevs.org/uploads/three-states-plan-ai-of-fear.pdf as seen in the classic shooter, F.E.A.R or HTN Hierarchical Task Networks https://en.wikipedia.org/wiki/Hierarchical task network . Those systems use a search algorithm to dynamically find a sequence of Actions which move the world into a Goal State. There’s nothing remotely so complex happening here ↩ fnref1 Animation ids are largely based on playstation controller inputs which are then offset by a per-actor value in the npc definition. Moveset swaps can be performed by changing the offset dynamically from scripts ↩ fnref2 I’m glossing over a small problem… You want to be able to write your scripts so that sub-goals function as a queue, not a stack, so they are executed in the order they’re pushed. Unfortunately that slightly complicates the implementation and explanation so I’ve left it as an exercise for the reader. ↩ fnref3 I’m not entirely sure whether they update only the current Goal from the top, or whether they recursively update currently active goals, but I suspect it might be the latter. This is still dramatically more efficient than re-evaluating decision criteria in a behavior tree. ↩ fnref4 Obviously this is a simplification, hopefully it’s clear I’m talking about the core data structure and framework rather than the large amount of peripheral work required to draw the rest of the owl. ↩ fnref5