What is known about robots and humanoids working safely next to people: the standards (IEC 61508, ISO 13849, ISO 10218), safety-rated sensors and components, who argues what, and the dates that decide it, with a source for every fact.
Functional safety is the part of safety that depends on a control system doing the right thing: a sensor that must see a person, a controller that must decide to stop, a drive that must limit speed. Its root standard is IEC 61508, edition 2.0 of 2010 with edition 3 in preparation (IEC webstore), and for machines ISO 13849-1:2023 with Performance Levels (ISO) and IEC 62061 with Safety Integrity Levels (IEC webstore). Physical AI pushes on it from two sides. The learned part is not covered: IEC 61508 "contains no usable information on addressing artificial intelligence in the context of functional safety" (KANBrief 2/22, IFA), and ISO/IEC TR 5469:2024 describes the "use of non-AI safety related functions to ensure safety for an AI controlled equipment" (ISO). And the hardware under the AI now has to arrive with safety evidence: Sony's IMX568 global-shutter image sensor and IMX560 SPAD depth sensor obtained IEC 61508 "SIL 2 Ready" certification in 2026 (Sony Semiconductor Solutions), SICK sells a 3D time-of-flight safety camera at PL c (SICK), and FORT Robotics' Trust Layer is certified to SIL 3 (The Robot Report, 18 Aug 2026). The rules are moving at the same time: ISO 10218-1 and -2 were revised in 2025 (ISO), the EU Machinery Regulation applies from 20 January 2027 and lists safety components with "self-evolving behaviour using machine learning approaches" among its high-risk products (EUR-Lex), and ISO 25785-1 for humanoids and other dynamically stable robots is a committee draft not expected before 2028 (ISO; Fraunhofer IPA, 27 May 2026). Humanoids are the hardest case: a humanoid needs power to stay upright, so the oldest safety move in robotics, cutting the power, can make it fall (A3, 23 Apr 2026).
Can a learned policy ever be certified, or only boxed in?
IEC 61508 "contains no usable information on addressing artificial intelligence in the context of functional safety", and systems that use AI inside a safety function are viewed as highly critical (Steimers and Bömer, IFA, KANBrief 2/22).
ISO/IEC TR 5469, "Functional safety and AI systems", was published on 8 January 2024; it covers AI inside a safety function, AI used to design safety functions, and the "use of non-AI safety related functions to ensure safety for an AI controlled equipment" (ISO). A technical report sets out guidance, not requirements that can be certified.
ISO/IEC TS 22440-1, the requirements for functional safety of AI systems, is a committee draft whose comment period closed on 11 April 2026; its joint working group (ISO/IEC SC 42 with IEC SC 65A) is convened by Riccardo Mariani of NVIDIA (ISO; IEC, 19 Oct 2023). NVIDIA's inspection lab checks Digit's safety-related software, AI components and cybersecurity against IEC 61508, ISO 13849 and ISO/IEC TR 5469 before third-party certification (NVIDIA, 22 Jun 2026).
What must a sensor or component vendor deliver?
IEC 61508 sets four Safety Integrity Levels by the rate of dangerous failures; for a continuously running safety function SIL 2 means between 10⁻⁷ and 10⁻⁶ dangerous failures per hour (itemis glossary). ISO 13849-1:2023 (edition 4, April 2023) rates safety-related parts of machine control systems against a required Performance Level (ISO); IEC 62061 is "a machinery sector specific standard within the framework of IEC 61508" (IEC webstore).
Sony Semiconductor Solutions: the IMX568 global-shutter image sensor "has obtained IEC 61508 SIL 2 Ready certification" (listed 26 May 2026), and so has the IMX560 SPAD time-of-flight depth sensor (listed 27 April 2026), which Sony pitches for detecting people and objects on AGVs and AMRs (Sony products news; Sony IMX560). "SIL 2 Ready" is a trademark of SGS TÜV Saar and means the component is approved by an independent accredited body for its probabilistic capability in SIL-rated safety applications (Sony global shutter page).
Sony's slide at ImagingNext in Munich on 15 October 2026 showed the IMX568 certificate, FS/71/220/26/2194, issued by SGS-TÜV Saar in Puchheim on 28 April 2026, tested to IEC 61508-1:2010 and IEC 61508-2:2010, and supporting camera systems for IEC 61508 SIL 2 and ISO 13849 PL d; a footnote said Sony did not apply the IEC 61508 development process but its ISO 9001 quality management system (Sony presentation, ImagingNext 2026).
What changes with ISO 10218:2025, ANSI/A3 R15.06-2025 and the EU Machinery Regulation in January 2027?
ISO 10218-1:2025 (industrial robots) and ISO 10218-2:2025 (robot applications and cells) were published in February 2025, replacing the 2011 editions (ISO, part 1; ISO, part 2). The US adoption, ANSI/A3 R15.06-2025, gives "Integrated guidance for collaborative robot applications, consolidating ISO/TS 15066" (A3); ISO's own page says part 2 complements ISO/TS 15066, which ISO has marked "to be revised", with a new ISO 15066-1 on biomechanical thresholds at an early stage (ISO, part 2; ISO/TS 15066).
ANSI/A3 R15.06-2025 recognises cybersecurity vulnerabilities as potential physical safety hazards (A3, 23 Apr 2026). ISA/IEC 62443 covers "electronically secure industrial automation and control systems" and bridges process safety and cybersecurity (ISA). Independent researchers found a root-level Bluetooth flaw and unannounced telemetry on the Unitree G1 (IEEE Spectrum; our note).
The EU Machinery Regulation (EU) 2023/1230 applies from 20 January 2027 after a corrigendum moved the date from 14 January (EUR-Lex). Its Annex I Part A, the products with the stricter conformity route, lists "Safety components with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions" and machinery with embedded systems of that kind (EUR-Lex). Its essential requirements say such control systems "shall not cause the machinery or related product to perform actions beyond its defined task and movement space", require safety decision data to be kept for a year, and require hardware to be "adequately protected against accidental or intentional corruption" (EUR-Lex).
How do you stop a humanoid that falls when you cut its power?
Machine stops come in three categories under IEC 60204-1: category 0 is "stopping by immediate removal of power to the machine actuators", category 1 is a controlled stop with power kept until the machine has stopped, category 2 a controlled stop with power left on (Machinery Safety 101). Drives provide these as certified functions such as safe torque off and safe stop 1 under IEC 61800-5-2 (IEC webstore).
A3 states that for a dynamically stable robot the usual emergency stop response, removing all power, is "NOT the optimal response" because the falling robot becomes a new hazard (A3, 23 Apr 2026). A humanoid study by Siemens researchers puts it the same way: "removing power from a walking biped causes an uncontrolled fall, so classical de-energization is itself a hazard" (Ding et al., arXiv 2608.02809).
The IEEE Humanoid Study Group proposes a five-level ladder for stop and power-loss behaviour, from an immediate power cut (level 0) to a robot that always holds a pose in which a power-loss collapse falls away from people (level 4) (IEEE study group, Sep 2025).
Can a humanoid fall without hurting anyone?
The IEEE study group says bipeds are always "one mistake or unexpectedly large disturbance away" from a hazardous fall, that "100% guaranteed stability is not always possible", and that standards for fall response are needed; it proposes a fall recovery test and, as an example, a stance that makes a power-loss collapse fall away from a nearby person (IEEE study group, Sep 2025).
A3 lists mechanical joint brakes to "prevent a full collapse" as the last layer of protection (A3, 23 Apr 2026).
SafeFall cut peak contact forces by 68.3% in falls of a Unitree G1, but its goal is protecting the robot's hardware, not bystanders (arXiv 2511.18509).
How hard may a robot touch a person?
Collaborative robots have two main modes: in speed and separation monitoring "no contact is allowed between the robot and human while the robot is moving" (Automation World, 23 Jul 2020); power and force limiting "imposes constraints on the energy absorbed during collisions prior to robot stopping" (Svarny et al., arXiv 1908.03046). ISO 10218-2:2025 brings "the full incorporation of the previous ISO/TS 15066" for collaborative applications, in SICK's reading (SICK, 8 Sep 2025).
For power and force limiting, the IEEE group says it is "completely unknown" what contact limits apply to humanoids, because ISO/TS 15066 and ISO 10218:2025 are written for single manipulators (IEEE study group, Sep 2025).
Fraunhofer IPA measured collision forces above 500 N on a Unitree G1, well over the pain thresholds the safety standards allow (Fraunhofer IPA, 27 May 2026).
Who certifies, who is liable, and what does it cost?
Certifiers are building physical AI services. NVIDIA's Halos AI Systems Inspection Lab is an "ANAB-accredited ISO/IEC 17020 Inspection Body" (NVIDIA, 22 Jun 2026) that prepares robots for certification by TÜV Rheinland, UL Solutions, TÜV SÜD, exida, SGS and CertX (Robotics 24/7, 22 Jun 2026); SGS and CertX joined the programme in August 2026 (GGBA, 12 Aug 2026).
First humanoid certificates are market-access marks, not proofs of safe AI behaviour: TÜV Rheinland certified AiMOGA's Mornine humanoid for CE, which the company calls a world first (TechNode Global, 26 Sep 2025), and AGIBOT's A3 Ultra "has received RED DA, RED NB and Safety AK certifications from TÜV Rheinland" (AGIBOT, 4 Sep 2026). For arms, TÜV SÜD certified Agile Robots' Yu 5 Industrial against IEC 61508, ISO 13849-1 and ISO 10218-1 (Agile Robots, 14 Feb 2024); NEURA states its MAiRA cobot is "Certified PL e / SIL 3 with TÜV-approved safety design" (NEURA Robotics).
SGS-TÜV Saar, which issued Sony's mark, is a joint venture of SGS Germany and TÜV Saarland with offices in Sulzbach and Puchheim near Munich, and also certifies microcontrollers to IEC 61508 SIL 2 (SGS-TÜV Saar).
When will standards for humanoids and mobile robots exist, and who writes them?
ISO/CD 25785-1 (TC 299) covers industrial mobile robots with actively controlled stability, legged or wheeled, in workplaces where the public is excluded; it was approved as a project on 22 May 2025 and its committee draft comment period closed on 8 July 2026 (ISO).
The working group, ISO TC 299 WG12, has about 120 participants from 17 nations, including all well-known humanoid makers from Europe, the US and Asia; final publication is planned for May 2028, with a first public draft somewhat earlier, and in Europe EN harmonisation follows (Jens Müller, Automationspraxis, 20 Feb 2026).
Federico Vicentini of Boston Dynamics is described as convenor (chair) of the ISO working group, ISO TC 299 WG 12 (MIT Technology Review, June 2025; SICK, December 2025; the ISO page names no convenor), with people from Agility involved; his aim: "We want to standardize the goal, not the way to get to the goal" (MIT Technology Review, 11 Jun 2025). Agility is a project leader for the US technical report ANSI/A3 TR R15.108 (Robotics 24/7, 15 Sep 2026).