Physical Interaction Through Non-Intervention: A Process-Regulation Framework for When Embodied AI Should Act, Defer, or Deliberately Remain Physically Inactive

Abstract

Embodied artificial intelligence is commonly evaluated by what it can perceive, plan, and do. Human-robot interaction research, however, increasingly shows that capability is not sufficient: an agent must also determine when action would interrupt a human, when assistance can be offered without disrupting an ongoing plan, when an instruction should be refused or clarified, and how interaction should adapt to the user and the history of the relationship. These problems are usually studied under separate labels such as interruptibility, non-intrusive assistance, abstention, expertise alignment, shared autonomy, mutual adaptation, and relational memory. This conceptual paper proposes a process-regulation view that treats these not as identical problems, but as neighboring manifestations of a common pre-action question: what transition should the human-AI system permit now? The proposed decision space contains three distinct outcomes: physical action, deliberate physical non-intervention, and defer/clarify. The framework is derived from longitudinal human-LLM work with AWM (AI Working Method) and conditionally transferred to embodied AI. Its central proposal is that the relative weight of human, agent, environmental, and historical signals should vary locally with the current goal, domain-relevant competence, preserved trace, safety constraints, and feedback from actual consequences (R_d). The paper connects this proposal to existing work and derives testable propositions and evaluation measures for future HRI experiments.