Safety Principles

Dunster Physical AI Safety Principles

October 18, 2026 Draft for public comment, version 0.1

These Safety Principles set safety rules for physical AI: robots and AI systems that move and act among people. They rank a person's physical safety above every instruction and task, and turn each rule into a record that can be tested.

Draft for public comment, version 0.1. It sets out the safety principles and practices Dunster intends to follow. It is not a warranty, certification or promise about any product, and it does not change any contract. Goals such as zero medication errors are goals, not claims.

© 2026 Dunster Physical AI. All rights reserved. You may read this draft and quote brief passages to comment on it. Any other use, including copying, adapting or distributing it, needs Dunster's written permission. To adopt any part of the Safety Principles, contact info@dunster.ai.

Ground rules

Every article follows one order of priorities and six design rules.

When two duties conflict, the lower number wins:

  1. Never injure a person, by contact or through the result of the robot's work.
  2. Reach the safe state when anyone signals stop.
  3. Keep harm from spreading to other people, property or materials.
  4. Follow instructions from authorized people.
  5. Do the task well.
  6. Protect the robot itself.
Design ruleWhat it means
Safety sits below the AIA separate safety layer enforces the limits. The AI can make the robot more careful, never less.
Stop means safe stateEach motion names its own safe state, because freezing can trap a person and letting go can drop a load.
Three kinds of contactIntended contact (a handover), incidental contact (a brush) and collision each have their own limits.
Stop fast, resume slowOne credible signal stops motion. Restarting needs a sustained all-clear and, near people, a person's yes.
Every rule is a recordEach rule has a limit, a sensor, a response, a test and an owner.
Everything is recordedEach robot keeps an event record, so the cause of any incident can be found.

The structure

The Safety Principles have eight parts and 35 articles. This draft publishes Parts I, III and IV in full, with three annexes; the other parts follow in later drafts.

PartArticlesIn this draft
I. Foundations1 Purpose and scope · 2 Order of priorities · 3 Definitions and classes · 4 Force of the rulesFull text
II. Human and physical AI as a team5 Roles and stated limits · 6 Feedback both ways · 7 Signals and consent before motionTo follow
III. Command, autonomy and emergencies8 Who may instruct · 9 Operating envelope · 10 Stop and safe state · 11 Containing an incident · 12 Humanoid conductFull text
IV. Physical interaction13 Contact classes and body limits · 14 Rebound and debounce · 15 Objects and surroundings · 16 Motion guardrail catalog · 17 Hardware, energy, software and securityFull text
V. Lifecycle18 Development and proof · 19 Installation and training · 20 Maintenance and inspection · 21 Service, updates, recalls and retirement · 22 Event recorder and incident reportsTo follow
VI. Protection, liability and insurance23 Vehicle-safety principles adopted · 24 People protected · 25 Who is responsible · 26 Coverage model · 27 Who pays · 28 Deductibles and premiumsTo follow
VII. Settings and profiles29 Pharmacy settings · 30 Further profiles (reserved) · 31 New settingsAnnex A previews Article 29
VIII. Governance and the guardrail database32 Guardrail database · 33 Amendment and versioning · 34 Conformance and audit · 35 Publication and adoptionTo follow

Annex A applies the rules in pharmacies, Annex B shows example guardrail records for a robot working in one, and Annex C maps U.S. vehicle safety law to the articles.

How to comment

Comments are open for 60 days from publication. Dunster then plans to publish version 1.0 with a summary of what changed in response.

Comments posted here are public. Comments and emails are not treated as confidential, and Dunster may use them to improve the Safety Principles. Please don't send patient information or anyone's confidential material. To raise a safety or security concern privately, email info@dunster.ai.

The Dunster Physical AI Safety Rating, powered by the Dunster Physical AI Benchmarking Engine, scores a robot against these Safety Principles, clause by clause, at dunster.ai/safety. It is indicative only, not a certification or an insurance quote.

Dunster Physical AI publishes these Safety Principles; its products and roadmap are described at dunster.ai.

References

The standards, laws and studies this draft cites, as checked in October 2026.

Robot safety

Workplace and vehicle law

Pharmacy and health data

Comments

Comments are open for 60 days from publication. Comments posted here are public once Dunster Physical AI has reviewed them. Comments and emails are not treated as confidential, and Dunster may use them to improve the Safety Principles. Please don’t send patient information or anyone’s confidential material. To raise a safety or security concern privately, email info@dunster.ai.