Abstract
Researchers from the Army Research Laboratory have introduced Hybrid Team Tetris (HTT), an open-source research platform designed to study complex human-AI coordination. By layering multi-agent learning and evolving rules onto the classic game, HTT provides a low-barrier environment for testing how humans and machines can collaborate in unpredictable, high-speed future scenarios.
Introduction
As AI capabilities expand rapidly, a "moving target" problem emerges: machine intelligence often evolves faster than our understanding of how humans should interact with it. Future environments for hybrid teams will likely be far more complex than today's, yet research is often hindered by the high barrier to entry. Studying hybrid intelligence typically requires deep expertise in both human-centric disciplines, such as psychology, and machine-centric ones, like reinforcement learning.
Hybrid Team Tetris (HTT) addresses this by providing a "shared vision"—a simplified, "toy" research platform that allows multidisciplinary teams to experiment with complex teaming concepts without needing to build bespoke simulations from scratch.
Methods and Architecture
The authors chose Tetris as the foundation for HTT due to its well-defined rules, computational accessibility, and long history as an AI benchmark. To transform this solo puzzle into a hybrid teaming platform, several key features were implemented:
- Deep TAMER (Teaching Agents Manually through Evaluative Reinforcement): This framework allows humans to train AI agents via simple reward signals (e.g., pressing "ENTER") rather than complex coding.
- Multi-Agent/Multi-Human Architecture: The platform supports configurations where multiple humans guide multiple agents. This includes "dependent agents" that learn from the collective decisions of the team.
- Advanced Gameplay Modes: To simulate future challenges, HTT includes hidden rules (revealed only to subsets of the team), "superhuman" scaling (managing up to 10 simultaneous games), and rapid rule changes (altering shapes or scoring mid-match).
Key Insights: Intelligence as a System Property
A core tenet of the HTT project is that in a hybrid team, individual entities do not necessarily need to be "intelligent" in isolation. Instead, the hybridization—the deliberate blending of processing between humans and machines—is what must demonstrate the ability to learn and reason.
Furthermore, the authors argue that studying how humans use current technology may be misleading. Just as GPS has altered human spatial memory, future AI will fundamentally change human behavior. HTT allows researchers to simulate these future "refined" human capabilities by pushing players into supervisory roles at superhuman speeds.
Results and Future Work
The primary outcome of this work is the functional release of the HTT platform. The authors successfully integrated the Deep TAMER framework into a multi-game environment, demonstrating that non-experts can train agents across multiple simultaneous boards.
A key proof-of-concept is "Agent 5," illustrated in the hierarchical architecture, which learns not from a human directly, but from the outputs of four other human-guided agents. This represents a "team-of-teams" learning structure that can be scaled for complex organizational research.
The platform is currently open-source on GitHub. Future work will focus on community contributions to increase the sophistication of human-machine interactions and cross-community collaboration to bridge the gap between disparate R&D silos.
Summary updated