# A methodology for integrating AI into embodied human intelligence for the performance of complex tasks

## TL;DR

This paper introduces a framework for combining human and AI intelligence to perform complex tasks, using a layered model and dynamic Bayesian networks to enhance human performance.

## Contribution

A novel four-layer dynamic Bayesian network framework that models human-AI collaboration for embodied tasks, enhancing performance through bidirectional inference and real-time adaptation.

## Key findings

- The framework achieved high agreement with clinicians in automated rehabilitation assessment (90.8-93.1%).
- Clinicians reported increased confidence and efficiency when using AI insights for therapy planning.

## Abstract

We propose a theory and methodology for designing human-artificial intelligence (AI) collaboration in complex, embodied tasks. The theory distinguishes human embodied intelligence from computational intelligence and identifies synergies in which AI enhances—rather than replicates or replaces—human performance. We represent observable structures of expert performance as a nested network with four interdependent layers: Environment (space and tools), Activity (what is done), Goals (what is aimed for), and Meaning (how performance is interpreted), all connected by dynamic four-layer edges. A bidirectional Dynamic Bayesian Network (DBN) computes this representation across temporal scales: instants, actions, complete performances, and sequences. The DBN informs the design of digital tools (from sensors to data structures and AI modules) that capture human performance and extract features, descriptors, and predictions that enhance the observability and analysis of performance. During task performance, a top-down pass predicts expert orientation—current goals and interpretations—and drives a search policy that selects where to look. A bottom-up pass processes action-conditioned computational observations and filters them through a gated pipeline to produce new candidates for four-layer connectivity (c4). After expert validation, candidates update the network, sharpening DBN posteriors, reducing entropy, and thereby enhancing human performance. We instantiated this framework in automated physical rehabilitation assessment through a 12-month deployment with 10 clinicians and 105 stroke survivors. Co-design cycles developed and enriched a four-layer DBN representation of rehabilitation assessment and informed the design of a computational ensemble for automated assessment. The computational ensemble achieved 90.8% agreement with clinicians at the exercise level, 93.1% at the segment level, and 90.6% at the movement quality level. Clinicians validated automated assessments at high rates and reported improved confidence and efficiency when leveraging ensemble insights for therapy assessment and planning. This portable methodology and theory can be applied to the embodied performance of complex tasks across multiple applications.