[ Deep Dive ]

Bibliography

The annotated sources behind every Deep Dive page. Each entry has a stable anchor (#ref-N) so research pages link back here as [N]. Found a paper or talk we should cite? Open an issue.

GOAP foundations → 2025 talks

If you find a paper or talk we should cite, open an issue — this bibliography is meant to grow.

GOAP foundations

[1] Orkin, J. (2006). Three States and a Plan: The A.I. of F.E.A.R. Game Developers Conference, San Francisco. The original GOAP-in-production paper. Established A* over symbolic world state as the foundation of F.E.A.R.'s AI but did not address replan stability beyond the planner's per-call correctness. PDF: web.media.mit.edu/~jorkin/gdc2006_orkin_jeff_fear.pdf.

[2] Jacopin, É. (2015). Optimizing Practical Planning for Game AI. In Rabin, S. (ed.), Game AI Pro 2: Collected Wisdom of Game AI Professionals, Chapter 13. CRC Press / A K Peters. Formally characterises GOAP's replan overhead and goal-oscillation problems using F.E.A.R.'s shipped levels as the dataset; identifies the "rats keep replanning" anti-pattern. Chapter PDF: gameaipro.com/GameAIPro2/GameAIPro2_Chapter13_Optimizing_Practical_Planning_for_Game_AI.pdf.

[3] Conway, C. (2015). Goal-Oriented Action Planning: Ten Years Old and No Fear! Game Developers Conference, San Francisco. A production retrospective ten years after Orkin's original talk. Reaffirms commitment / stability as the persistent open problem in production GOAP. GDC Vault: gdcvault.com/play/1022019.

AAA postmortems

[4] Girard, S. (2021). AI Action Planning on Assassin's Creed Odyssey and Immortals Fenyx Rising. Game Developers Conference. Ubisoft production talk on shipping GOAP at AAA scale, including the per-agent stability work needed to keep open-world NPCs behaving coherently.

[5] Boeda, G. (2023). Let Your Agents Plan Together: Multi-Agent Cooperation With GOAP. AI Summit, Game Developers Conference. Square Enix talk on extending GOAP to multi-agent coordination — the territory beyond per-agent stability.

[6] Marko, M. (2025). Combining GOAP and MBTs to Create NPCs' Behaviors for 'Kingdom Come: Deliverance II'. Game AI Summit, Game Developers Conference. Warhorse Studios production talk demonstrating GOAP remains a live AAA technique in 2025, paired with Markov Behavior Trees as the layered control framing.

[13] Hanlon, S., Watts, C. (2017). Behavior Decision System: Dragon Age Inquisition's Utility Scoring Architecture. In Rabin, S. (ed.), Game AI Pro 3, Chapter 31. CRC Press. BioWare production talk on the utility-scoring architecture they shipped, including the commitment / hysteresis layer they added when scores hovered near equal. Another AAA case alongside Odyssey [4], FFXV, and KCD2 [6].

Cross-paradigm evidence (the same problem at different layers)

[10] Dawe, M. (2015). Preventing Animation Twinning Using a Simple Blackboard. In Rabin, S. (ed.), Game AI Pro 2, Chapter 6. CRC Press. "Animation twinning" — multiple agents picking the same animation off the same trigger — is flapping at the animation layer. Same problem, different layer; cross-paradigm evidence that stability must be addressed wherever oscillation can emerge.

[12] Francis, A. (2017). Overcoming Pitfalls in Behavior Tree Design. In Rabin, S. (ed.), Game AI Pro 3, Chapter 9. CRC Press. Catalogues BT re-entry, transition thrash, and decorator oscillation patterns. BT's flavor of flap, demonstrating that the problem is paradigm-general and not unique to GOAP.

Architectural framings

[9] Côté, C. (2014). Reactivity and Deliberation in Decision-Making Systems. In Rabin, S. (ed.), Game AI Pro: Collected Wisdom of Game AI Professionals, Chapter 11. CRC Press. Frames the axis on which our toolkit operates: reactivity (respond fast to world change) vs deliberation (commit to chosen plan). Each anti-flap family is a different point on that axis. Chapter PDF: gameaipro.com/GameAIPro/GameAIPro_Chapter11_Reactivity_and_Deliberation_in_Decision-Making_Systems.pdf.

[11] Dill, K. (2015). Dual-Utility Reasoning. In Rabin, S. (ed.), Game AI Pro 2, Chapter 3. CRC Press. Named anti-oscillation technique: pair a "raw utility" with a "smoothed utility" and use the smoothed value for selection. Directly maps to our Family 4 (EMA-smoothed scalar inputs to a Family-1 score-domain selector).

Theoretical context

[7] Pereira, R. F., Pereira, A. G., & Meneguzzi, F. (2019). Using Sub-Optimal Plan Detection to Identify Commitment Abandonment in Discrete Environments. arXiv:1904.11737. Frames goal commitment in symbolic planners as a detectable property of plan trajectories, giving the broader theoretical context for "should the agent abandon this plan?" arxiv.org/abs/1904.11737.

Implementation references

[8] crashkonijn (2026). GOAP — A multi-threaded GOAP system for Unity. Open-source plugin documentation. The most-active production-grade GOAP plugin in any engine today (v3.1.2, ~1.7k GitHub stars, 863 commits as of March 2026); ships goal re-evaluation throttling at the selection layer. The Unity counterpart to Intent Forge's UE5 niche. github.com/crashkonijn/GOAP · goap.crashkonijn.com.

Recent practitioner record (2024–2025)

[14] Jacopin, É. (2025). AI Planning Analytics — From F.E.A.R. (2005) to Assassin's Creed: Shadows (2025). AI and Games Conference, London. A twenty-year longitudinal study of planner analytics across two flagship shipped titles. The strongest single data point that GOAP-derived planning is a live AAA technique in 2025, not a historical artifact. Supplements the earlier Game AI Pro 2 chapter [2]. aiandgames.com — Conference 2025 announce.

[15] Smrček, P. (2025). Supporting Thousands of Simulated NPCs in the Open World of KCD2. AI and Games Conference, London. A second Warhorse production talk alongside Marko's GOAP+MBT talk [6] — focused on scale of deployment rather than control architecture. Evidence that the KCD2 GOAP-derived system runs at open-world NPC counts in shipping form. Linked from Orkin's conference recap [16].

[16] Orkin, J. (2025). One Trillion Parameters and No Plans. AI and Games Conference, London. Twenty-year-anniversary talk by the creator of GOAP, presented opposite the 2006 F.E.A.R. paper [1]. Argues that symbolic planners (GOAP, HTN) remain necessary in the post-LLM era because LLMs cannot replace planning at runtime, and advocates a neurosymbolic future where LLMs streamline content authoring while planners handle execution. Live demo of the Bitpart.AI prototype illustrating the pattern. Conference recap on LinkedIn · Talk on YouTube via AI & Games.

[17] Guerrilla Games / Decima Engine team (2024). HTN Planning in the Decima Engine. AI and Games Conference 2024. HTN-in-production talk from the studio behind Horizon Zero Dawn and Horizon Forbidden West. Companion HTN data point to the GOAP record cited above. Cited on the HTN vs GOAP page. YouTube — AI and Games Conference 2024.

Community / community-tier sources

[18] Aversa, D. (2018, updated 2020). Choosing between Behavior Tree and GOAP (Planning). davideaversa.it. The canonical community essay on when to choose GOAP over Behavior Trees. Lays out the two motivating criteria (action-space combinatorics, designer's inability to enumerate transitions ahead of time) and the three recurring drawbacks (higher implementation complexity, runtime cost, loss of designer control over emergent behavior). Influences the framing used on our When to use / When not to use and Positioning pages. davideaversa.it/blog/choosing-behavior-tree-goap-planning.