Abstract
In high-stakes, time-critical environments such as healthcare and disaster response, team performance often degrades due to misaligned mental models and partial observability. This paper presents a framework for an "AI Coach" designed to improve human teamwork by inferring unobservable intents through imitation learning. By modeling teamwork as a latent variable problem, the system detects coordination failures and provides real-time, cost-aware interventions to realign the team. Our approach moves beyond rigid, hand-coded rules by learning from past demonstrations, offering a scalable solution for enhancing collaboration in complex, dynamic domains.
Introduction
In high-pressure scenarios, effective teamwork is frequently undermined by "partial observability." Team members often lack a complete view of their environment and their colleagues' actions, and the cognitive or temporal cost of communicating every detail is prohibitively high. This leads to the breakdown of "shared mental models"—the common understanding of the team's goals, plans, and individual roles. When these models diverge, coordination fails, even if individual members are performing their tasks correctly in isolation.
While human coaches are effective at identifying and correcting these misalignments, they are a scarce resource. Existing automated solutions often rely on expert-defined, hand-coded rules that are difficult to scale and adapt to new environments. This research proposes an automated AI coaching system that learns the nuances of "good teamwork" by observing past human demonstrations and intervenes during live tasks to ensure the team remains "on the same page."
Methodology
The proposed system consists of three integrated components designed to model, detect, and correct coordination failures.
1. Bayesian Team Imitation Learner (BTIL)
The BTIL serves as the core analytical engine. Unlike traditional imitation learning, which focuses solely on mapping observable states ($s$) to actions ($a$), BTIL explicitly models unobservable intents ($x$) as latent variables. Using mean-field variational inference, the system learns intent-aware policies (what an agent does given their goal) and intent dynamics (how goals evolve over time). This Bayesian approach is specifically designed to be sample-efficient, a critical requirement when working with limited human demonstration data.
2. Detection of Imperfect Coordination
Once intents are inferred, the system assesses the degree of coordination by measuring the compatibility of these latent variables across the team. If the AI determines that team members are pursuing incompatible goals or possess conflicting understandings of the current task state, it identifies a mental model misalignment. This allows the system to flag potential failures before they manifest as physical errors.
3. Task-time Intervention for Collaboration (TIC)
The TIC component manages the delivery of advice. It treats intervention as an optimization problem, balancing the benefit of realignment (preventing errors) against the cost of intervention (the potential to distract or overwhelm the team). Because finding an optimal intervention schedule is computationally intensive, the system employs heuristic strategies that analyze the distribution of possible intents to provide the most effective guidance at the most opportune moment.
Key Insights and Ideas
- Intent as a Latent Variable: A central premise of this work is that actions alone are insufficient for understanding teamwork. By modeling the hidden "intent" behind actions, the AI can distinguish between a team that is coincidentally aligned and one that is truly coordinated.
- Proactive Alignment: By using shared mental models as a primary metric, the AI coach can intervene preventatively. This is a shift from reactive systems that only provide feedback after a performance drop is detected.
- The Cost of Help: A critical insight is that excessive intervention can be counterproductive. An effective AI coach must be "intervention-aware," recognizing that every piece of advice carries a cognitive load that could hinder execution in time-sensitive tasks.
- Non-Adversarial Scaling: While adversarial methods like Generative Adversarial Imitation Learning (GAIL) are popular, they can be unstable in multi-agent settings. This research advocates for non-adversarial Bayesian approaches combined with function approximators to scale to high-dimensional, real-world data.
Results and Discussion
Preliminary evaluations indicate that the BTIL algorithm is highly efficient, learning complex team behaviors without requiring massive, perfectly labeled datasets. This efficiency is vital for practical deployment in specialized fields like emergency medicine.
Safety is a primary motivator for this work. Previous iterations (Seo et al. 2021) demonstrated that intent misalignment in simulated high-stakes environments could lead to "fatal accidents." The current system addresses this by providing a mechanism for real-time realignment. Furthermore, the research introduces a method for "inference under intervention," which allows the AI to continue tracking human intents even as its own advice changes the humans' decision-making processes.
Limitations and Future Work
While promising, the system currently faces challenges in scaling to very high-dimensional or continuous state spaces. Current efforts are focused on integrating deep function approximators to enhance the BTIL’s capacity. Additionally, the TIC system currently assumes a pre-defined set of possible intents. Future research aims to develop "indirect interventions"—such as interpretable advice or environmental cues—that do not require an exhaustive list of every possible human goal. Finally, while human demonstration data has been used for training, the full closed-loop validation of the AI intervening with human teams in real-time remains an ongoing area of experimental focus.
Summary updated