Skip to main content
Back to timeline
arXivSource publication:

A team can be modeled as a ball in an energy bowl: ANCHOR forecasts four-person drift with 52 readable parameters

Synopsis

Collaborative mixed-reality headsets already record who speaks, who looks at the same image and who stands close, but prompting a team before it drifts needs a plant model that learns from few teams, stays bounded when iterated and can be read by a designer. ANCHOR treats a four-person team as a ball in an energy bowl whose bottom blends recent behavior, the team's own history and other teams' norms, written as a switched port-Hamiltonian system with 52 parameters. On 17 held-out four-person teams, about 24% of a deviation remains after 32 s against a predicted 30%; conversation-collapse warning reaches AUROC 0.85, and accuracy matches a black-box baseline with about 40,000 parameters.

Source-provided article image: ANCHOR: A Port-Hamiltonian Model for Forecasting Team Dynamics in Human-in-the-Loop Systems
arXiv

Interpretation

When four people sort images together in collaborative mixed reality, their interaction state can be written as a dynamical system with an explicit energy, and a deviation from the usual pattern fades the way a ball rolls back to the bottom of a bowl. Earlier forecasts of team interaction relied on black-box sequence models or simple statistical rules, with no small, iterable, readable plant model; team psychology offered only qualitative descriptions, not a computable form of return. On 17 four-person teams held out from training, after a speaking burst the speaking share falls from 46% to 27% 32 s later against a forecast of 30%; over two minutes the forecast error is 14.1 points, below 17.0 for the session mean and 22.6 for persistence.

The model writes a team state as 20 numbers: four for member speaking time, six for each pair's shared gaze, six for each pair's close co-presence, and four team-level quantities including speaking balance. Team-level quantities are kept as separate coordinates rather than derived from member forecasts; a member is silent 44% of the time, and balance derived from the member forecast has 3.7 times the error of forecasting it directly (2.86 against 0.77). Removing each part and retraining on held-out teams raises member-and-pair error by 19.3% without the blended resting point (better on 17/17 teams), by 11.3% without the dynamics (17/17), and raises team-level error by 270.3% without the team-level state (16/17).

The bottom of the bowl is not fixed: at every forecast origin it is a blend of the last five windows, the session so far and the training teams' average. Using only recent behavior or only session history as the resting point is worse; the three sources get different learned weights by behavior, with shared attention almost entirely on the team's own history (0.81) and speaking balance following the moment and other teams (0.43 and 0.56). A recent-behavior-only resting point raises member-and-pair error by 19.3% and team error by 16.5%; a session-history-only resting point raises them by 3.4% and 9.7%, both significant across the 17 held-out teams.

Because the matrices are built from team structure, energy balance and boundedness under spaced cues hold for every fitted parameter value, with no post-fit stability check. A free linear map can only be checked for stability after each fit, whereas ANCHOR's positive-definite regulation and skew-symmetric exchange let forecasts only shrink toward the bowl bottom without task input. In 60-window (16-minute) rollouts for all held-out teams with phase switching and task input, the energy balance held at 100% of steps and every forecast coordinate stayed between 0 and 1.

Perspective

The results come from 17 four-person peer teams in one room sorting 28 images into six emotion categories by discussion in collaborative mixed reality, each wearing a Meta Quest Pro, with no time limit and no scripted turns; the fitted anchors and gains describe this task, and what transfers is the model structure, which needs refitting for a new task. The cue-timing comparison is done in simulation with assumed cue effects swept over a range, so the effect of real prompts has not been measured.

Phase, the task port, sensed damping and inertia are on par with the reduced model on held-out teams, and exchange gains are small, so at this resolution teams slide back to rest rather than swing; the phase switches once in 15 of 17 sessions, typically about 5 minutes in, so only 11% of forecasts meet a switch within 48 s. Per-second analysis found no more time-irreversibility or rotation in real teams than in surrogate teams, but the conversation signal could not be settled because of cross-talk between co-located headsets. Cues are absent from the recorded sessions, so the simulation reflects only how well a policy times cues for a given effect.