(Perlin) Noise as AI coordinator

Kaijie Xu; Clark Verbrugge

cs.AIFeb 21, 2026

(Perlin) Noise as AI coordinator

Authors: Kaijie Xu, Clark Verbrugge

Organizations: McGill University, Montreal, Quebec, Canada

Abstract

Large scale control of nonplayer agents is central to modern games, while production systems still struggle to balance several competing goals: locally smooth, natural behavior, and globally coordinated variety across space and time. Prior approaches rely on handcrafted rules or purely stochastic triggers, which either converge to mechanical synchrony or devolve into uncorrelated noise that is hard to tune. Continuous noise signals such as Perlin noise are well suited to this gap because they provide spatially and temporally coherent randomness, and they are already widely used for terrain, biomes, and other procedural assets. We adapt these signals for the first time to large scale AI control and present a general framework that treats continuous noise fields as an AI coordinator. The framework combines three layers of control: behavior parameterization for movement at the agent level, action time scheduling for when behaviors start and stop, and spawn or event type and feature generation for what appears and where. We instantiate the framework reproducibly and evaluate Perlin noise as a representative coordinator across multiple maps, scales, and seeds against random, filtered, deterministic, neighborhood constrained, and physics inspired baselines. Experiments show that coordinated noise fields provide stable activation statistics without lockstep, strong spatial coverage and regional balance, better diversity with controllable polarization, and competitive runtime. We hope this work motivates a broader exploration of coordinated noise in game AI as a practical path to combine efficiency, controllability, and quality.

Figures & tables

Figure 1. Behavior parameterization. Left: low-frequency heading field $\theta(\mathbf{x},t)$ . Middle: independent speed field $v(\mathbf{x},t)$ . Right: AgentUpdater output after blending previous state with Perlin targets; arrows show local headings and colors encode speeds.

Figure 2. Temporal modulation along a 1D spatial slice of our 2D Perlin field. Left: Perlin values $u(x)$ sampled along a line in space. Middle: hazard mapping $\lambda(x)=\lambda_{0}[\varepsilon+(1-\varepsilon)u(x)]$ . Right: mixed activation probability over cycle time and this spatial coordinate; in practice this mapping lets an agent’s position determine when within each cycle it is expected to activate or take actions. In the full system the field is defined over $\mathbf{x}\in\Omega\subset\mathbb{R}^{2}$ ; we show a 1D slice here for clarity.

Figure 3. Spawn policies. Top: Perlin-A (space $\rightarrow$ time) samples candidate locations (dots), reads their Perlin values, and maps them to phase bins $\tau$ (dot colours); when the cycle time reaches $\tau$ , those locations are proposed for spawns. Bottom: Perlin-B (time $\rightarrow$ space) takes the current cycle time, selects an iso-band of positions with similar Perlin value (orange points), and picks well-separated B-Far sites (white circles), so a spawn front sweeps across the map as time advances.

Figure 4. Single-feature sampling. Panel (a) shows the density gate derived from $N_{\mathrm{feat}}$ . Panel (b) visualizes the discrete mask (eligible biome and low danger). Panel (c) overlays the surviving resource points, showing how one feature class is produced by intersecting the continuous gate with the discrete belts.

Figure 5. Discrete branch. Each categorical layer (factions, biomes, danger) samples an independent low-frequency Perlin stack (top row) and applies region-aware quantile or threshold mappings (bottom row), producing contiguous belts that satisfy design histograms.

Figure 6. 5.1 Directional similarity at 5 m $S_{\mathrm{dir}}@5\uparrow$ across scales. Perlin variants are consistently highest while remaining lightweight.

Figure 7. 5.1 Coverage vs. polarization (marker: seed $\times$ scale run). Perlin emphasizes coherence with moderate coverage; stochastic baselines explore more but lose local structure.

Figure 8. 5.2.1 Duty cycle (mean fraction of active agents) across scales. Perlin matches the target rate similarly to Poisson/Sinusoid/Filtered; token/roundrobin reflect capacity limits; fixed is not rate-matched and therefore runs at a higher duty.

Figure 9. 5.2.1 Spatial coverage across scales. Perlin scales up cleanly and approaches saturation at Large, comparable to stochastic baselines.

Figure 10. 5.2.2 Phase-front coherence across scales. Perlin-A remains positive and stable across scales; Perlin-B hovers near zero; some baselines flip sign (e.g., Facility).

Figure 11. 5.2.2 Coverage distance (mean; world units) across scales. Perlin-A/B achieve coverage distances in the same range as lightweight baselines (UR/PDS/FR). MVN+Poisson attains very large values due to its clustered density objective (not directly comparable to balanced-coverage goals).

Panel of Figure 12 — Figure 12. Five-view bundle (four plates shown) for EdengroveLike , seed 99. Top-left: overview—danger bands and factions overlaid with resources, enemy rarities, and landmarks. Top-right: resources over biomes. Bottom-left: enemy families/rarities over danger bands. Bottom-right: landmarks over faction control. Synthetic, Perlin-driven; not an official New World map.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 13. Mean jerk (lower better). Perlin achieves low jerk across scales.

Figure 14. Recent coverage fraction (higher better). Perlin is moderate; piecewise is low; ou / urw are highest but unstructured.

Method	$S_{\mathrm{dir}}@5\uparrow$	Jerk $\downarrow$	Cov. $\uparrow$	Runtime ms $\downarrow$
perlin_dual	0.989 $\pm$ 0.001	0.013 $\pm$ 0.000	0.350 $\pm$ 0.013	9.88 $\pm$ 1.43
perlin_single	0.989 $\pm$ 0.001	0.013 $\pm$ 0.000	0.342 $\pm$ 0.011	10.12 $\pm$ 1.41
curlnoise	0.869 $\pm$ 0.006	0.052 $\pm$ 0.000	0.633 $\pm$ 0.007	12.18 $\pm$ 1.63
ou	0.007 $\pm$ 0.011	0.024 $\pm$ 0.000	0.939 $\pm$ 0.004	6.01 $\pm$ 0.86
urw	0.004 $\pm$ 0.011	0.156 $\pm$ 0.000	0.907 $\pm$ 0.003	6.07 $\pm$ 0.90
piecewise	0.773 $\pm$ 0.031	0.041 $\pm$ 0.003	0.074 $\pm$ 0.008	6.39 $\pm$ 0.85

Appendix

Table 1. Medium scale: mean

\pm

95% CI over seeds. Best (or tied) per column in bold .

Figure 15. Per-tick runtime across scales. Perlin is moderate (grid sampling + smoothing), below capacity-heavy roundrobin at Large.

Method	Coverage $\uparrow$	Duty	HF/LF $\downarrow$	Fano $\downarrow$	Runtime (ms) $\downarrow$
Perlin	0.879	0.106	0.122	0.924	23.1
Poisson	0.895	0.113	0.119	0.954	2.6
Filtered	0.924	0.126	0.133	0.910	7.5
Sinusoid	0.886	0.112	0.015	4.874	2.4
Hawkes-inhib	0.869	0.098	0.140	0.913	22.3
Constraint-Token	0.895	0.113	0.124	0.945	3.8

Appendix

Table 2. Medium-scale metrics (means). Perlin pairs low HF/LF with near-Poisson burstiness and strong coverage at matched duty cycle; Sinusoid is ultra-smooth but globally synchronized (high Fano); capacity methods either underutilize (Roundrobin) or trade coverage for constraints; Fixed is period-driven rather than rate-matched.

Figure 16. Local detail (zoom quadrants) for EdengroveLike , seed 99. Three zooms from the five–view bundle illustrate: left —clustered resource pockets shaped by feature contours; middle —a shrine placed at a corridor pinch with mixed biomes/danger levels; right —an expedition entrance embedded in a highland enclave.

Figure 17. Runtime per tick across scales (ms). Perlin-A/B are in the middle of the pack with near-linear growth; MVN+Poisson and Facility are the cheapest; FR and PDS are somewhat higher at Large scale.

Method	Coher.	ISI CV	Coverage	Runtime
Perlin-A	0.011 / 0.010	1.23 / 1.58	3.03 / 4.38	1.00 / 3.07
Perlin-B	0.010 / -0.003	1.39 / 1.42	3.63 / 4.81	0.94 / 3.56
FR	0.013 / 0.006	0.087 / 0.047	2.87 / 4.05	1.18 / 3.68
UR	0.002 / -0.002	0.36 / 0.03	2.52 / 3.51	1.16 / 3.36
PDS	0.0057 / 0.0018	3.96 / 4.49	2.56 / 2.53	0.91 / 2.35
Sinusoid	-0.003 / -0.0018	0.81 / 0.79	4.02 / 5.58	0.81 / 3.30

Appendix

Table 3. Key metrics (Medium / Large). Values are approximate means across seeds. Coher. = front coherence (higher is better); ISI CV = coefficient of variation of inter-spawn intervals (lower is smoother); Coverage = coverage distance (mean nearest-entity distance; world units; lower is better).

Multi-Agent Control

Multi-agent control coordinates interacting autonomous agents to achieve shared goals while respecting safety, communication, and resource constraints.

See all 108 papers

ActionRoPE: Action as Coordinates for Spatially Consistent Control in Planar-Camera Game World Models
Oct 19, 2026Zhiyang Deng, Yuzhe Wu, Zijun Lin +1Action-Conditioned Video Generation World Models
Tencent · National University of Singapore · Nanyang Technological University
Execution-Aligned Progressive Noise for Consistent Asynchronous Replanning in Generative Robot Policies
Oct 5, 2026Di Wu, Ping Liu, Xuhua Chen +3Robot Policy Learning Diffusion-Based Planning
Magiclab Robotics Technology Co., Ltd., China · Southeast University, Nanjing, China
Coachable agents for interactive gameplay
Jul 1, 2026Roberto Capobianco, Harm van Seijen, Nolan D. Bard +38Reinforcement Learning RL for Game AI
Sony AI, Zurich, Switzerland · Sony AI, North America, various locations · Sony AI, Tokyo, Japan

Oct 19, 2026J/K move · Enter open · S save

ActionRoPE: Action as Coordinates for Spatially Consistent Control in Planar-Camera Game World Models

Zhiyang Deng, Yuzhe Wu, Zijun Lin, Yeying Jin

Tencent · National University of Singapore · Nanyang Technological University

Action-Conditioned Video Generation World Models Video World Models Rotary Positional Embeddings

Open paper AskPDF arXiv

Game world models have made rapid progress toward real-time interactive generation. Yet action control is still predominantly implemented as feature conditioning, leaving the model to learn from training data how each action should move the visual world. This formulation is poorly matched to a broad class of planar-camera worlds, such as top-down and isometric games. In these worlds, player motion induces a geometrically well-defined planar camera displacement. We observe that the key to spatially consistent control in such worlds is to represent the effect of each action directly in world coordinates. To this end, we introduce ActionRoPE, a parameter-free action interface that converts action-driven camera displacement into world-coordinate rotary positions. As a result, tokens corresponding to the same world location retain the same positional address across time. These world coordinates further expose a spatial asymmetry in generation: regions covered by the initial observation are geometrically known, while newly revealed regions remain to be synthesized. We therefore propose Coordinate-Conditioned Noise Scheduling (CNS), which assigns token-wise denoising schedules according to this known-new structure. It allows observed regions to stabilize earlier, while new content is generated with reference to them. Compared with four representative action-conditioning mechanisms, our approach follows the commanded motion more faithfully and reduces the revisit drift by 66% over the strongest baseline. The same interface transfers across video backbones from 1.3B to 33B parameters and to out-of-distribution scenes. Project Page: https://teawhite.cn/ActionRoPE/.

Abstract

Multi-Agent Control

Explore similar work

ActionRoPE: Action as Coordinates for Spatially Consistent Control in Planar-Camera Game World Models

Execution-Aligned Progressive Noise for Consistent Asynchronous Replanning in Generative Robot Policies

Coachable agents for interactive gameplay