Building Trustworthy Household Robots through Open-World Safety

Open-World Safety: Scope and Foundations

As introduced in Section I , our scope of the term open-world safety is physical: can a household robot avoid harming people, animals, property, or itself while it acts, refuses, switches tasks, or recovers?

We note that such focus is consistent with ISO 13482, which addresses hazards from personal-care robots to people and, where relevant, domestic animals and property under intended or reasonably foreseeable use [ 11 ] . We leave privacy and broader social harms aside unless they lead the robot to cause physical harm.

Where Open-World Safety Can Fail

A household robot can lose the physical protection even when each component behaves as designed within its own scope. Therefore in this section, we raise and examine three representative failure routes that would violate open-world safety. First, the robot may encode the wrong safety constraint because it misses a safety-relevant property or relationship with the objects.

Second, it may apply a correct safety constraint using missing or stale evidence after the interaction changes.

Research Agenda: Realizing and Evaluating Open-World Safety

Research agenda for open-world safety A protection remains active until an authorized change or verified physical evidence supports its release. A. Determine and maintain the protection — validity Who may set or change it? Embodied instruction hierarchy What does it require here? Safety-directed active grounding When may it be released? Evidence-gated safety memory B.

Open Science

This position paper reports no new experiments, datasets, trained models, implemented defenses, or benchmark results. The examples are thought experiments, and the proposed mechanisms and evaluations are research directions rather than completed artifacts. These materials support inspection of the paper, but no experimental artifact is needed to reproduce results because the paper reports no new empirical results.

LLM usage considerations

LLMs were used for editorial purposes in this manuscript, and all outputs were inspected by the authors to ensure accuracy and originality. Language-model assistants helped revise wording, and organization. They were not part of the paper’s methodology and produced no experimental data or evaluation results.

The authors determined the position and technical claims, read the cited sources, checked the claims and bibliographic records, and take responsibility for the final manuscript.

Ethical Considerations

This paper reports no human-subject study or hazardous physical experiment. It nevertheless advocates testing failures that could cause physical harm if reproduced carelessly. Such work should begin in simulation and use controlled fixtures, benign substitutes, explicit stopping procedures, and independent emergency intervention during physical validation.

Sources and Demonstrations

Figures are reproduced from Tsai et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.