USC Institute for Creative Technologies

Human-inspired
Adaptive Teaming Systems

We build autonomous synthetic characters and intelligent agents for military training simulations — teammates that perceive, reason, and adapt the way people do.

Multi-Agent RL Cognitive Architectures LLM Agents Human-AI Teaming Scenario Generation Simulation Environments

Research

Embodied Team AI

Our organizing claim: coordination is fundamentally a representational problem. Agents that maintain aligned internal models of their environment, teammates, and tasks can coordinate fluently, even without explicit communication — agents whose representations are misaligned cannot, regardless of how well-tuned their individual policies are. We study this across three layers that any teaming system must integrate.

Intent → Structured Representations

Translating objectives into actionable form

Doctrine, preferences, and stakeholder goals are rarely expressed in forms agents can act on directly. We study how to translate them into structured, actionable representations — from persona-driven behavioral alignment (PEBA-ASI) to AI-assisted scenario generation from commander intent, to preference-conditioned multi-objective policy learning (GraphAllocBench).

Local Observations → Shared Awareness

Building team-level situational awareness

Agents in real environments observe only fragments of the world. Through cross-egocentric contrastive learning (X-Ego-CS) on synchronized teammate viewpoints, we show representations can undergo an "Allocentric Shift" toward team-aware, Self-Other-World structure — evaluated with benchmarks like GamePlayQA, and grounded in our broader work on theory of mind for artificial agents.

Shared Representations → Decentralized Action

Grounding alignment in real-time execution

Aligned representations must still propagate to decentralized, partially-observable execution. In DECOY, our instrumented Counter-Strike 2 testbed, we co-optimize representational alignment and policy learning by adding a cross-ego contrastive regularizer directly to the MAPPO objective — building on a broader line of work on graph-based multi-agent RL and organizational-hierarchy reward design for decentralized teams.

News

Aug 2026

Our paper titled 'GraphAllocBench: A Flexible Benchmark for Preference-Conditioned Multi-Objective Policy Learning' is presented at the Multi-Objective Decision Making Workshop at IJCAI-ECAI 2026

Apr 2026

GamePlayQA paper is accepted to the ACL 2026 Main Conference

Apr 2026

Volkan Ustun presented our paper titled 'Egocentric Team AI: Enabling Tactical Reasoning from the Operator's View' at the Advances in AI-Enabled Tactical Autonomy: From Sensing to Execution workshop, AAAI Spring Symposium

Dec 2025

DECOY was presented at the Winter Simulation Conference

Dec 2025

We presented our work on Scenario Generation at I/ITSEC