ANCHOR: A Port-Hamiltonian Model for Forecasting Team Dynamics in Human-in-the-Loop Systems

Diana Romero, Salma ElmalakiarXiv 2026 · arXiv

TL;DR

ANCHOR is a physics-inspired model that treats human team dynamics as a "ball in an energy bowl," using a port-Hamiltonian framework to forecast collaboration behavior. By mapping team psychology principles to just 52 interpretable parameters, it matches the accuracy of black-box models with 40,000 parameters while providing formal guarantees for stability in human-in-the-loop systems.

What Problem Does This Paper Address?

In high-stakes environments like operating rooms or mission control, team collaboration is critical. However, teams often drift into "breakdowns"—situations where one person dominates the conversation or the group loses its shared focus. While Mixed Reality (MR) headsets can now sense who is speaking and where people are looking, we lack a reliable "plant model" to turn this data into actionable interventions.

Existing machine learning models for human behavior often fall into two traps:

  1. The Data Hungry Trap: They require thousands of examples, but recording high-quality team interactions is expensive and rare.
  2. The Black-Box Trap: They are uninterpretable and can produce "runaway" unstable predictions when iterated in a closed loop.

The authors seek a model that is small enough to learn from a few dozen teams, readable by designers, and mathematically guaranteed to stay bounded when used to trigger real-time coaching prompts.

Methods

The authors introduce ANCHOR, a model based on Port-Hamiltonian (pH) systems—a framework typically used to model physical systems like robots or electrical circuits.

The "Ball in a Bowl" Analogy

ANCHOR models a team's state (a 20-dimensional vector including speaking time, gaze, and proximity) as a ball in an energy bowl.

  • The Energy ($H$): Represents how far a team has drifted from its "resting point."
  • The Resting Point ($q^*$): The bottom of the bowl, calculated as a blend of three "anchors": the team's immediate behavior, its session history, and general population norms.
  • Regulation ($R$): The "friction" that pulls the team back to its resting point.
  • Task Input ($gu$): External forces (like making progress on a task) that move the bowl itself.

Key Innovations

  • Psychology-to-Physics Mapping: Each of the 52 parameters corresponds to a testable claim from team psychology (e.g., "teams restore balance after a disturbance").
  • Passivity by Construction: Because it uses the pH framework, the model is "passive." It is mathematically impossible for the model to predict infinite energy growth, making it safe for closed-loop controllers.
  • Permutation Equivariance: The model treats all team members as peers; renaming "Member A" to "Member B" doesn't change the forecast.

Datasets

The study utilizes the MURMR dataset, which consists of:

  • 17 co-located teams of four people.
  • Task: Sorting 28 images into emotion categories via consensus in Mixed Reality.
  • Sensors: Meta Quest Pro headsets capturing audio, head pose, and gaze.
  • Metrics: 20 values per 32-second window, including individual speaking shares, pair-wise shared attention, and "speaking balance" (entropy of talk time).

Key Insights and Ideas

  • The Power of Constraints: By forcing the model to obey physical laws (energy conservation and dissipation), the authors achieved high accuracy with orders of magnitude fewer parameters than a standard neural network.
  • The Three-Anchor Blend: A team's "normal" state isn't static. The most accurate forecasts came from blending what the team just did with their long-term habits and general human social norms.
  • Interpretability as a Feature: You can look at the fitted $R$ matrix and see exactly how many seconds it takes for a "speaking burst" to fade (typically 20–60 seconds). This allows designers to "read" the team's social dynamics directly from the model weights.
  • Timing is Everything: In simulations, using ANCHOR to time prompts (e.g., inviting a quiet member to speak only when a breakdown is forecast) was 2.2x more efficient than simply prompting on a fixed schedule.

Key Results

  • Accuracy: ANCHOR (52 parameters) matched the performance of a multi-layer perceptron (MLP) and a Ridge regression baseline with ~40,000 parameters.
  • Predicting Collapse: The model anticipates "conversation collapse" (when one person takes over) with an AUROC of 0.85.
  • Natural Decay: Real-world team deviations fade quickly; after 32 seconds, only 24% of a deviation remains. ANCHOR predicted 30%, showing it captures the natural "self-regulating" rhythm of human groups.
  • Data Efficiency: ANCHOR remains robust even when trained on as few as four teams, whereas black-box models' performance degrades significantly with small datasets.

Limitations and Future Work

  • Simulated Actuation: While the sensing and forecasting were tested on real team data, the effect of the prompts (the "cues") was only evaluated in simulation. A real-world deployment study is needed to see if humans actually follow the "nudges."
  • Task Specificity: The current anchors and gains are fitted to an image-sorting task. It remains to be seen how well these parameters generalize to different types of work (e.g., high-stress emergency response).
  • Temporal Resolution: The model operates on 16-second strides. The authors found no evidence of "lossless exchange" (energy moving between behaviors) at this scale, suggesting that finer-grained interactions might require even higher-resolution data.

For more details, see the full paper: "ANCHOR: A Port-Hamiltonian Model for Forecasting Team Dynamics in Human-in-the-Loop Systems."

Summary updated

Abstract

Cyber-physical-human systems close a loop around people by sensing human behavior, predict it and act on it. Human teams are a natural target, because collaboration breaks down in ways that a timely prompt could prevent. Collaborative mixed reality makes this loop practical, since each headset senses who speaks, where its wearer looks and who stands together, and it can deliver a light prompt. In this loop, the headsets are sensors, the prompts are actuators and the team is the plant. Closing the loop needs a plant model that is small enough to learn from few teams, safe to iterate and readable by a designer. We propose ANCHOR, which models a team as a ball in an energy bowl. The bottom of the bowl is where the team rests, and it is set by the team's recent behavior, its own history and other teams. Formally, ANCHOR is a switched port-Hamiltonian system with 52 parameters, and each part stands for a testable claim from team psychology. Because of this structure, for any fitted parameters, a forecast moves away from rest only when task progress or a prompt pushes it, and prompts of bounded size that are spaced in time move it a bounded distance. On \nGroups{} four-person teams held out from training, a team that drifts from its usual behavior returns most of the way within half a minute. After 32 s, about 24% of the deviation remains, and ANCHOR predicted 30%. Forecasts are clearly better when the resting point blends its three sources than when it uses one. ANCHOR matches the accuracy of a black-box baseline with about 40,000 parameters. It anticipates conversation collapse with an AUROC of 0.85. In simulation under assumed prompt effects, it removes 2.2 times as much breakdown time per prompt as periodic prompting.
BibTeX
@misc{romero2026anchor,
  author = {Diana Romero and Salma Elmalaki},
  title = {{ANCHOR: A Port-Hamiltonian Model for Forecasting Team Dynamics in Human-in-the-Loop Systems}},
  year = {2026},
  eprint = {2611.19124},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19124},
}
Figure