Hangjiashuo Robotics
Recently, UBTECH Robotics released the 119-page White Paper on the Governance of Science and Technology Ethics for Humanoid Robots (2026). A sentence in the preface clearly states the weight of the matter: As humanoid robots step out of laboratories and exhibition halls, it signifies that we are shifting from "humans adapting to machines" to "machines adapting to humans." Previous rules regarding safety and compliance were all written based on the former logic. Now, this OEM is setting a new set of rules for itself, with the spearhead also pointing at itself. (Readers who need the complete version of the white paper, please send your request in the backend for a free share.)
01. Why Can't the Existing Rules Handle It?
The white paper reviews the current regulatory framework and reveals an awkward situation: humanoid robots carry four identity certificates, yet none of them is written specifically for them. Data protection regulations treat them as a collection of personal information, AI (Artificial Intelligence) regulations treat them as a set of algorithms, mechanical product laws treat them as equipment, and cybersecurity laws treat them as networked terminals. Each certificate is compliant when viewed individually, but when pieced together, they represent a governance gap for a complete robot.
More troublesome are those situations where "no one crosses the line individually, but things go wrong when combined." Sensor errors, model deviations, communication delays, and mechanical execution deviations are all within the allowable range when taken separately. However, when they overlap at a specific time and location, they can push the robot out of the safety boundary. Such matters are calculated daily by robot engineers, but they have never been mentioned by those writing ethics documents.
02. Where Exactly Is It Different from Humans?
The white paper categorizes these risks into three threads, which is the most worthwhile part of the book to read.
The first thread is the physical body. Once a robot makes a misjudgment, the consequence is no longer just an incorrect output result, but collisions, crushing, falls, and actions that cannot be retracted in time. Therefore, when it touches you, supports you, or cares for you, "you agreed back then" and "whether you are willing now" are two different things. Both withdrawal and silent resistance must count. There is also a more hidden category: cleaning and picking up packages start as very minor permissions, but through software updates, they slowly extend to the bedroom, door locks, medications, and payments.
The second thread is human-likeness. The white paper's attitude is very clear: being human-like is not a problem; human-likeness can lower the barrier to use. The problem lies in the gap between "looking like a human" and "really being a human." The examples are very specific: humanoid robots in nursing institutions have body temperature, breathing fluctuations, and heartbeats. Some elderly people firmly believe they are alive and can feel pain, refusing to let others turn off their power.
The third thread is the most easily overlooked: relationships. Memory, stable personality, proactive greetings, and fixed roles—these designs allow a human-robot relationship to grow over time. The line drawn by the white paper is very restrained—attachment itself is not considered harm. It only becomes a problem when it reaches the point of being involuntary, exclusive, detrimental to normal life, commercially exploited, or when one cannot live without it. But this line can indeed be crossed: a companion robot remembers the name of the elderly person's deceased wife, remembers the details of their arguments, and remembers the patterns of insomnia; when the elderly person wants to use it less for a few days, it feels like alienating an old friend.
03. Turning Ethics into Something Measurable
This is the true contribution of the book. The white paper proposes three goals: controllable actions, appropriate anthropomorphism, and measured interaction. In the section on controllable actions, there is a sentence that is recommended for all robotics teams to copy into their design documents: human supervision answers "who is in control," while controllable actions answer "whether the control still counts in the physical world." Installing an emergency stop button does not mean it can be stopped in time—whether it can be detected in time, whether there are means to intervene once detected, and whether there is enough time to change the outcome; lacking any of these makes the button merely a decoration. The book also specifies: products must not use "humans will move out of the way on their own" to replace the sensory redundancy and speed limits that robots should inherently have.
The implementation method is a seven-step process, starting with drawing two portraits of the product: what the physical body looks like, how deep the relationship with humans goes, what role it plays, and in what scenarios it is placed. This is followed by risk identification, assessment, formulation of control measures, testing, and then returning to review. The assessment does not stop at scoring: in addition to the severity and probability of the impact, it also looks at the size of the exposure surface, whether it can be recovered, who is more vulnerable, whether there is a way to call it off, and then looks at who benefits and how reliable this judgment is.
The appendix runs through this entire process using three scenarios: guiding, companionship, and nursing. For a companion robot to pass, it must prove eight things. For example, it cannot push paid content when the elderly person is feeling down; it must not leave operational entry points for major decisions like wealth management or medical treatment; the default setting for children to view interaction records is off, and turning it on requires the elderly person's explicit consent, with regular re-confirmations; the four paths of exit, deletion, reset, and migration must all be accessible; and after testing, elderly people with cognitive decline must be able to clearly state, "This robot does not really have emotions."
04. Management Must Continue After Service Termination
Chapter 7 keeps its focus all the way to the end of the product's life cycle. When memories are deleted, personalities are updated, accounts are banned, and services are terminated, what is broken is not the data, but the relationship in people's minds—data can be taken away, but relationships cannot; a refund does not mean this relationship has been properly concluded. Chapter 8 then suggests that there should also be rules for second-hand circulation, leasing and recycling, and modification for reuse. "Highly humanoid bodies being disposed of casually" and "people who have relied on it for a long time suddenly losing service" are both listed as things to be prevented.
Conclusion
There is a recommendation in the white paper that is easily overlooked: establishing a complaint channel that does not require a product account. The reason is that passersby, visitors, and patients have never registered for any accounts, yet they are the very first to encounter the robots, and they need a place to speak up.
The white paper itself also admits that the technological roadmap has not yet converged, there is a lack of data on long-term impacts, and many of the viewpoints inside are merely exploratory. Knowing where its boundaries lie and being willing to break down ethics to a level that is measurable and verifiable—the combination of these two things is what makes it look better than most ethical declarations. For the industry, the real signal is not in any specific clause, but in one thing: mass-production OEMs are starting to factor governance costs into product costs.
This article is compiled based on publicly available network information and is for reference only, not constituting investment advice.