Every day we edge closer to having robots in our homes. The usefulness of a typically bipedal humanoid robot for aiding in the household and providing an extra helping hand in daily life is theoretically unparalleled, but the safety of such systems is in its infancy. In this paper, we highlight some practical safety-related questions that system developers should consider throughout the design of a consumer robot. Some of these questions have straightforward guidance in safety-related standards; others are far more complex and require research into novel solutions.

Stability
Q1: What is stability, what are its related hazards, and how is it recovered?
Stability, in robotics, generally refers to the ability to maintain the center of mass of the robot within its support polygon both with (dynamic stability) and without (static stability) active control. Loss of stability can lead to hazardous situations such as falling and flailing which become hazardous when the robot makes physical contact with people, pets, or objects in the environment. While full prevention of stability loss would be the best solution, it is not reasonable; every dynamically stable robot will fall at some point. When looking at the safety of balance recovery, the speeds at which the robot’s joints can move is an important focus because of the direct correlation to force exerted on the environment as well as the available reaction time of anyone nearby. A balance must be struck between joint speed, recovery capability, and accepting inevitable falls.

ISO 25785-1, currently under development, will be a crucial guide for developers of Dynamically Stable Industrial Mobile Robots (DSIMRs) in defining safety requirements for stability. DSIMRs, per the standard’s abstract, include robots with an unspecified number of legs (e.g., bipeds, three-legged, quadrupeds) and wheeled balancing robots. What delineates these robots from those covered by existing standards is their “actively controlled stability,” which refers to the active control needed for balance and the potential for instability in the absence of power. While the standard excludes non-industrial robots, it is expected to set the precedent for evaluating stability strategies and metrics, and Reynolds & Moore is helping to shape its scope and direction. Evaluating the existing strategies, however, can still benefit developers seeking to optimize system cost, define stability metrics, and account for varying use cases. Zhang et al. (2025), for example, discuss the various technical implementations of fall and balance recovery in “A Review of Fall Coping Strategies for Humanoid Robots.” The authors’ consideration of fall detection, preventive actions against falls, and post-fall protection measures for robots across a wide variety of form factors promotes safe design practices while ISO 25785-1 continues to mature.

Motion Control
Q2: How is robot motion controlled to facilitate its safe interaction?
A primary goal of consumer humanoids is to perform tasks around the home, some of which will require exchanging objects with humans or other forms of direct interaction. Given the force output capability of a robot, these interactions could quickly become hazardous. Limiting this force output when in proximity to a human is a strong method for reducing the likelihood of a hazardous event occurring and can be done through joint velocity limiting or joint torque limiting. While a risk assessment can identify the need for these measures, their implementation can spur debates within developer teams. What constitutes a “safe value” for velocity or torque? These definitions could lead to tradeoffs in the robot’s interaction capabilities (i.e. how fast or with what level of strength it can perform a task), affecting the robot’s marketability. Further, designing in the sensing capabilities required to adequately monitor or estimate force output to SIL 2 or higher becomes an intricate balance when working within the space constraints of a humanoid form factor and the time constraints of real-time autonomous work.

ISO/TS 15066, which provides safety requirements for collaborative industrial robots, includes informative tables (Annex A) on contact force limits for different parts of the human body and should be considered when designing limits into a robot system. Developers can especially benefit from this guidance in defining solutions for elder care applications. For example, Miyagawa et al. (2020) propose using speed-based and contact force-based safety devices in “Consideration of Safety Management When Using Pepper, a Humanoid Robot for Care of Older Adults.” ISO/TS 15066 does clarify, however, that annex values are conservative estimates, which has led to criticism that achieving compliance means settling for inefficient robot performance. Gabrielli and Sechi (2021) as well as Bricher and Muller (2025) acknowledge that artificial intelligence (AI), which we will discuss next, can provide a path to reducing cycle time without compromising safety.

Intelligence & Hazard Recognition
Q3: What role does AI play in the robot’s behavior?
With the recent advances in AI, system reliability is playing both defense (cybersecurity) and offense (new capabilities). Humanoid robots are being marketed as having AI systems that allow them to seamlessly perform their tasks, but this comes with risk. Determining the possible adverse effects of model implementation first requires an understanding of the subsystem architecture (perception, control, interfaces, etc.) in which the model is utilized, then unraveling its effect on behavior. Are the models pre-trained or continuously learning? Pre-trained models can be validated as safe before reaching the public, but continuously learning models can spontaneously develop new capabilities with hazardous side effects. Does the robot use AI for object classification to add nuance to the manipulation of pieces of broken glass, or other hazardous objects, so as not to drop them or cut someone? Can the robot reject the use of an object such as an obvious weapon, and how robust is this rejection given that many objects could be ambiguously classified (e.g. baseball bats, pencils, broken glass bottles)? These questions are a bit difficult to answer, but we ask system developers to consider the high severity consequences that can result from a robot wielding a hazardous object and whether metrics evaluating model performance adequately account for them.

ISO/IEC CD TS 22440 (Artificial intelligence — Functional safety and AI systems) is an upcoming standard that will define safety requirements for AI-enabled systems that can adapt over time. The requirements will include a structured methodology to analyze, assess, and mitigate AI-related risks for industrial applications which, in the absence of other application-specific standards, may also provide useful guidance for non-industrial environments. Reynolds & Moore is helping to shape the scope and direction of this standard as well, while our own Director of Engineering, Paul Schmitt, co-authors The ML FMEA in Action: Lessons from Applications of Machine Learning Safety, a paper that bridges Machine Learning and Failure Modes and Effects Analysis (ML FMEA) and presents an analysis framework that can be applied today. Artur Gunia (2024) also recommends a systematic analysis of consumer use cases and behavior in “Safety of Household Robotics – ethical doubts,” calling out the need to consider inappropriate commands, interaction limits, gripping context, and multi-factor behavior models. Simultaneously accounting for human, robot, and object movement is a challenging problem, but Nie et al. (2018) provide an example multi-factor model, omitting robot movement, in “A child caring robot for the dangerous behavior detection based on the object recognition and human action recognition.” Testing these models and certifying them to relevant standards, a service provided by RMAI Safety Certification, not only supports the development of a safe AI-enabled system but also strengthens consumer trust in the robot’s ability to respond in unique scenarios, such as the recognition of hazardous objects.

Conclusion
Safety needs to be of the utmost priority when developing a robot intended for consumer home use. Research literature, such as the articles we referenced above, can build upon existing or inspire novel solutions to technical challenges. There are many standards currently in existence or in development that should also be leveraged. ISO 13482, which defines safety requirements for personal care robots, applies to all the considerations in this paper (hence, not discussed for any one question above) and should be a primary design reference. While awaiting that standard’s next release, consider ISO 10218, which serves as an equivalent for industrial robots and establishes a foundation of concepts that can be leveraged outside of the industrial environment. Of course, the IEC 61508 series, which generally defines the functional safety process, provides higher-level lifecycle and design guidance. Other upcoming standards, such as ISO 25785-1 for industrial DSIMRs and ISO/IEC 22440 for AI in functional safety, will define new requirements that shape the future of humanoid robot design. Reynolds & Moore brings unique value in having representatives on these technical committees and by hosting training for both of these developing standards. For development teams navigating any of these questions, Reynolds & Moore offers years of experience helping clients realize innovation safely in the robotics space, and we would welcome the chance to help.