CEAA: A Cognitive Embodied Agents Architecture for Interactive Computing Systems
Aimilios Hadjiliasi, Louis Nisiotis
cs.AI
2026-08-11
CEAA folds Sense-Think-Act and BDI into a reusable three-layer, twelve-component architecture for embodied virtual agents, still conceptual with only a 92-person educational study.
Deploying an embodied Intelligent Virtual Agent (IVA) with real cognitive ability in an interactive 3D environment is stuck between two poles. Low-level reactive control (state machines, behavior trees) is locked to commercial game engines and struggles with goal-directed reasoning, explainability, and personalization. High-level cognitive architectures (BDI, SOAR, ACT-R) are theoretically sound but too complex to fit in real-time 3D worlds. The authors (two UK-based researchers) want a practical middle layer: a cognitive architecture template that unifies high-level reasoning, memory, planning, and real-time embodied action, and that can be implemented directly.
CEAA (Cognitive Embodied Agents Architecture) has three layers and twelve components.
The User and Environment Layer holds users, agents, and virtual objects. The Knowledge Layer keeps a real-time log of environment state on a shared blackboard, recording events as shared problems agents can selectively address. The Agent Layer handles human-like reasoning and decision-making.
Among the twelve components: Sense (selective perception via rule filters and pandemonium-style algorithms), Memory (similarity-based and case-based recall), Think (central coordinator integrating memory and orchestrating modules), Cognitive Construct (the BDI triple of Belief, Desire, Intention), Reasoner (rule inference and constraint satisfaction to select intentions from desires), Planner (turning intentions into action sequences), Act (triggering animation, speech, navigation), and Behavior Mapper (mapping symbolic actions to concrete behavior specs). The architecture stitches Sense-Think-Act and BDI into a layered engineering template.
This paper runs no technical benchmark. The only empirical work is a 92-person mixed-methods educational study: 46 in a desktop condition and 46 in VR, doing a roughly 45-minute interactive task with pre- and post-tests. Both groups showed significant within-group learning gains, with the desktop group slightly higher than VR but no significant between-group difference. Both rated the system as relatively easy to use and useful; VR scored slightly higher on effort expectancy and behavioral intention.
| Dimension | Result |
| Learning gains | Significant within-group in both; no significant between-group difference |
| Technology acceptance | Both rated easy to use and useful; VR slightly higher on effort expectancy and behavioral intention |
| Technical performance | Not evaluated |
To be clear: the study measures usability of students learning with the system, not the architecture's reasoning quality, latency, or scalability. Feasibility rests on a prototype demo, with no quantitative performance data.
For game AI or VR teaching-agent developers, CEAA works as a structural checklist: it lays out which modules an IVA needs (perception, memory, reasoning, planning, action) and how they exchange data, so one can build from it rather than translating BDI or SOAR theory. It explicitly targets Unity and Unreal, with extension tools promised as future work.
The cold water: the value of synthesizing existing paradigms into a new architecture depends on whether it is less effort than behavior trees in a real project, and the paper provides no such evidence.
The authors are candid: the architecture remains at a conceptual and theoretical level and needs benchmarking; it currently serves as reference and guidance, with Unity or Unreal tools left to future work; performance, scalability, and generalizability are untested. Reading it, one more concern: many of the twelve components (Memory, Reasoner, Planner) defer implementation details to future work, so today it reads more as a design document than a runnable system.