When the Fence Came Down Before the Standard Came Out: Agility's Digit 5 Through an ISO 25785-1 and ISO 10218-1:2025 Lens

There is no incident here. Nobody was hurt, no recall was filed, no regulator opened a docket. What happened on September 15 is that a company with a real deployment record and a real order book announced a 129 kg bipedal robot designed to share aisles with people and no longer stand behind a fence — and the only international standard written for that exact class of machine is still a committee draft. I have spent a lot of these posts reading failures backwards into the missing work product. This one is the rarer case where you can read the missing work product forwards, before the failure, which is the only time the reading is actually useful.

Agility Robotics is not a demo company. Digit has been on customer floors for close to three years, has logged more than 65,000 hours of operational time across GXO, Schaeffler, Amazon and Toyota Motor Manufacturing Canada, and hit a cumulative 100,000-tote milestone at GXO's Flowery Branch facility at roughly 98% on-task accuracy. That record is exactly why the Digit 5 announcement is worth a careful read rather than a press-release skim. The company is not asking for permission to experiment. It is telling a market with more than $300 million in multi-year orders behind it that the physical barrier is coming out of the risk assessment.

The public record

What was announced. On September 15, 2026, Agility Robotics unveiled Digit 5, describing it as the company's first humanoid "engineered for cooperatively safe work at scale." The machine stands 5 ft 11 in (1.81 m), weighs 284 lb (129 kg), reaches to 7.2 ft (2.2 m), and repeatedly lifts loads up to 50 lb (22.7 kg) — a payload Agility frames as full coverage of single-person lift tasks in OSHA-regulated facilities. A 90-minute battery charges in 9 minutes, a 10:1 run-to-charge ratio against Digit 4's 2:1, enabling more than 20 productive hours in a 24-hour day.

The safety architecture, as described. Agility says Digit 5 "continuously monitors for people using proprietary AI algorithms and multiple sensor technologies," and responds by avoiding the person, stopping, or assuming a seated position. Visual and auditory cues convey motion intent. Separately — and this is the sentence that matters — "an independent safety controller supervises Digit's response whenever someone is detected within an unsafe distance and instantly triggers the appropriate safety response." Agility is also the first launch partner for NVIDIA's Halos for Robotics platform, pairing with IGX Thor and Halos Core.

The standards position. Agility states it is a project leader and active contributor to ANSI/A3 TR R15.108, a technical report under development in the US and Canada for dynamically stable industrial mobile robots including humanoids, and to ISO 25785-1, which it describes as the first international safety standard for the humanoid category. It also claims Digit was the first humanoid to pass an independent field evaluation on a customer production line for industrial safety standards administered by OSHA.

Where ISO 25785-1 actually is. Per the ISO catalogue entry, ISO/CD 25785-1 — Robotics — Safety requirements for dynamically stable industrial mobile robots (legged, wheeled, or other forms of locomotion) — Part 1: Robots — is under ISO/TC 299. New project approved 2025-05-22. Committee draft registered 2026-05-08. CD consultation initiated 2026-05-12. Close of the CD comment period 2026-07-08. Current stage: 30.60. It has not been registered as a DIS. Part 2, covering integration of applications, is to be developed separately and does not yet exist.

Timeline. Digit 5 early access is expected in the first half of 2027, general availability by end of 2027, with a CE mark and additional marks planned for EU, Great Britain and Northern Ireland market access. So the product ships into the gap, and it ships into Europe.

The standards lens

The interesting thing about this machine is not that it is unregulated. It is that it falls cleanly between three published Type C standards, each of which excludes it for a specific, defensible technical reason.

ISO 10218-1:2025 / ISO 10218-2:2025 (US national adoption ANSI/A3 R15.06-2025, Parts 1 and 2 approved 21 August 2025, full three-part standard published 29 October 2025) is the industrial robot standard. Edition 3 is a genuine improvement: it folds the ISO/TS 15066 collaborative content directly into the body, retires "collaborative robot" as a property of the machine, makes functional safety requirements explicit rather than implied, replaces "safety-rated monitored stop" with "monitored standstill," and adds cybersecurity into safety planning. It is also written around a manipulator whose base does not move and whose stability is a structural property, not a control outcome.

ISO 3691-4:2023 and ANSI/A3 R15.08-1/-2/-3 cover industrial mobile robots. They give you active protective fields, stopping-distance-as-a-function-of-speed, restricted zones, fleet supervision. They assume the robot is statically stable — that if you cut power mid-travel, the machine decelerates and stays upright.

ISO 13482 covers personal care robots and explicitly scopes out industrial environments.

ISO/CD 25785-1 exists precisely because of the gap, and its own scope statement names it: the standard applies to robots with "actively controlled stability," which it defines as "a robot that requires an active control in order to remain balanced and could become unstable in the absence of power." Read that clause as a safety engineer rather than a standards editor and it says something uncomfortable.

For a dynamically stable robot, de-energization is not a safe state.

Every reflex the industrial safety community has built over forty years runs the other way. ISO 13850 Stop Category 0 — immediate removal of power to the machine actuators — is the default safe state of the machine world. It is what an E-stop mushroom does. It is what a safety relay does when the light curtain breaks. It is what STO means in a drive. On a 129 kg biped standing 1.81 m tall, Category 0 is not a mitigation; it is the hazard. Cutting power converts a controlled machine into 129 kg of unguided falling mass in a space you just removed the fence from.

This is why "assuming a seated position" appears in Agility's list of safety responses next to "stopping" and "avoiding." That is not a UX nicety. It is the safe state. And a safe state that requires energy, actuator authority and a working balance controller to reach inverts the ordinary dependency: the power path and the control path become part of the safety function rather than things the safety function switches off. A Category 1 controlled stop, or a Category 2 stop with monitored standstill, has to be the design point, and the reserve energy needed to complete the descent has to be a monitored, verified quantity — not a best-effort remainder.

The second structural issue is the perception split. Agility describes two distinct things: proprietary AI algorithms doing continuous human monitoring across multiple sensors, and an independent safety controller that supervises and triggers. That split is the correct architecture, and it is the correct architecture for a reason worth stating plainly: a learned perception model is not a safety-related part of a control system under ISO 13849-1. You cannot assign it a Performance Level. You cannot compute a PFH_D for it. It has no meaningful diagnostic coverage in the ISO 13849-1 Annex E sense, its failure modes are not systematic in a way Annex G's common-cause scoring anticipates, and its "false negative" rate is a statistical property of a data distribution rather than a component reliability figure.

So the question any integrator, notified body or plant safety manager should be asking about Digit 5 — and about every humanoid that follows it — is not "does it detect people." It is: which channel is the safety function, and what does the AI channel do when it is wrong in the direction that matters? If the AI channel is the performance layer (smooth avoidance, high availability, good throughput) and the rated channel is a conventional safety-rated sensing envelope that triggers the controlled-seat regardless of what the model thinks, that is a defensible architecture and a familiar one. If the rated controller merely actuates a decision the unrated model made, the PL of the combined function is the PL of the model, which is to say undefined.

A worked snippet

Below is my own worked risk assessment for the top hazard, done the way ISO 12100 §5.5 and ISO 13849-1 §4.3 would have you do it for a machine with no applicable Type C standard. This is an illustrative analysis of the published machine parameters, not Agility's analysis, and the PL assignments are mine.

Hazard identification and PL_r determination (ISO 12100 / ISO 13849-1 Annex A)

| ID | Hazardous event | Severity (S) | Frequency (F) | Avoidance (P) | PL_r | |---|---|---|---|---|---| | H-01 | Uncontrolled fall onto person in shared aisle following balance-control loss | S2 (irreversible, head/torso) | F2 (continuous shared space) | P2 (fall onset under 400 ms, no avoidance) | e | | H-02 | Transient contact with 22.7 kg payload during transfer motion | S2 (crush/impact, payload at 2.2 m reach) | F2 | P2 | e | | H-03 | Quasi-static clamping of limb between robot torso and rack upright | S2 | F1 (aisle-dependent) | P2 | d | | H-04 | Failure to reach seated safe state due to insufficient reserve energy | S2 (degenerates to H-01) | F2 | P2 | e | | H-05 | False-negative human detection in the performance (AI) channel | S2 (degenerates to H-01/H-02) | F2 | P2 | e |

Why the fall hazard cannot be argued down with biomechanical limits

ISO 10218-2:2025 carries forward the ISO/TS 15066 biomechanical tables. Those quasi-static and transient limits sit in the range of roughly 110 to 220 N and fractions of a joule to a few joules of transferable energy per body region. The fall case is not in the same regime:

Fall energy estimate, H-01 (order-of-magnitude, author's calculation)

  m        = 129 kg            (published mass, Digit 5)
  h_cg     = 1.0 m             (assumed standing CG height; not published)
  g        = 9.81 m/s^2

  E_pot    = m * g * h_cg
           = 129 * 9.81 * 1.0
           = 1,265 J           (upper bound, full CG drop, no dissipation)

  Reference: ISO/TS 15066 Annex A transferable-energy limits
             per body region are order 1 J.

  Ratio    ~ 3 orders of magnitude above the collaborative limit.

The conclusion is not subtle: you cannot make a fall safe by limiting power and force. Power-and-force-limiting is a mitigation for contact during controlled motion. For a dynamically stable biped, the fall is a separate hazard requiring a separate, higher-integrity mitigation — which is why fall-zone exclusion appears in the ISO 25785-1 work. Reported figures from the ISO 25785-1 drafting discussion put the fall zone of a 1.7 m robot travelling 1.5 m/s at roughly 2 m in any direction. Digit 5 is taller than that reference machine.

Fault tree, top event H-01

                    H-01  Uncontrolled fall onto person
                          in shared aisle
                                |
                              [ OR ]
            ______________________|______________________
           |                      |                      |
    Balance control          Safe state not         Person inside
    authority lost           reached in time        fall zone at
           |                      |                 onset
        [ OR ]                 [ OR ]                    |
     ______|______          _____|_____               [ AND ]
    |             |        |           |            _____|_____
  Actuator    Estimator   Reserve    Seat          |           |
  power       divergence  energy     trajectory  Detection   Motion
  loss        (IMU /      below      blocked by  channel     intent
  (BE-1)      contact     descent    obstacle    failed to   cue not
              fusion)     demand     (BE-6)      trigger     perceived
              (BE-3)      (BE-5)                 (BE-7)      (BE-8)
        |
   Undetected
   fault in
   power path
   (BE-2)

Note what the tree shows. The two branches on the left are entirely internal to the robot and are classical functional-safety territory — you diagnose them, you compute them, you put them under a rated architecture. The right-hand branch is the one that has no settled method, because BE-7 is a perception failure and BE-8 is a human-factors failure, and their AND gate is the whole case for keeping a physically rated detection envelope underneath the AI layer.

Derived requirements (excerpt)

Five traceable requirements from the above. IDs are mine; they are written the way I would expect to see them in a safety concept handed to an integrator.

| ID | Requirement | Traces to | Verification | |---|---|---|---| | HSR-DS-001 | The robot shall reach a stable seated or braced configuration within 1,500 ms of a safe-state demand, from any commanded posture at rated load, without exceeding the declared fall-zone radius. | H-01, H-04 | Demonstration on 100 demands across posture/load matrix; worst-case time recorded | | HSR-DS-002 | Reserve stored energy sufficient for the HSR-DS-001 descent shall be continuously monitored; the robot shall initiate the safe state and inhibit new task assignment when reserve falls to no more than 120% of the calculated descent demand. | H-04 | Analysis plus fault-injection on energy monitor; PL d per ISO 13849-1 | | HSR-DS-003 | Detection of a person inside the declared fall zone shall be implemented by a safety-rated sensing channel independent of the AI perception channel, achieving PL d, Category 3 per ISO 13849-1, with the AI channel restricted to performance-level avoidance only. | H-05, BE-7 | Architecture review; PFH-D calculation; validation per ISO 13849-2 | | HSR-DS-004 | A Stop Category 0 command from any source, including the ISO 13850 emergency stop device, shall not remove actuator power until the HSR-DS-001 descent has completed or 2,000 ms has elapsed, whichever is earlier. | H-01, BE-1 | Test on all E-stop actuation paths; documented deviation rationale per ISO 13850 §4.1.5 | | HSR-DS-005 | Declared travel speed shall be limited such that the resulting fall zone remains inside the safety-rated detection envelope of HSR-DS-003 with a margin of at least 300 mm, under worst-case floor friction in the declared operating environment. | H-01, H-02 | Stopping and fall-zone measurement per intended ISO 25785-1 method; re-run per site |

HSR-DS-004 is the one that will generate an argument, and it should. Delaying power removal after an E-stop press is heresy in a fenced cell and correct engineering on a biped. It is also exactly the kind of deviation that a Type C standard is supposed to bless so that every manufacturer does not have to litigate it alone with a notified body. ISO 25785-1 is at stage 30.60. Until it reaches publication, every one of these arguments is a bilateral negotiation.

What the headline really tells us

The headline says a humanoid robot no longer needs a cage. Read as an engineering statement, it says something narrower and more demanding: the protective function that used to be delivered by a steel fence — a Category B-by-geometry device with essentially infinite reliability and zero diagnostic requirement — has been moved into a software-and-sensor architecture that now has to carry the whole of that risk reduction on a computed Performance Level.

That is a legitimate trade. It is the same trade the automotive industry made when it moved from mechanical linkages to drive-by-wire and had to invent ISO 26262 to justify it. But the fence had one property the new architecture has to earn rather than inherit: it did not have a false-negative rate.

The missing artifact here is not a recall notice or an incident report. It is a published Type C standard — and, downstream of it, the Part 2 integration document that does not yet exist. Agility is doing the right thing by co-authoring both; being in the room at ISO/TC 299 and on ANSI/A3 TR R15.108 is materially better than shipping and waiting. But the ordering is still inverted. The machines go to early access in the first half of 2027 into a market including the EU and UK. ISO/CD 25785-1 has not left committee. Whatever a plant safety manager signs in 2027, they will be signing a risk assessment built on Type A and Type B standards written for machines that stay upright when you turn them off.

Ask the vendor the H-05 question, in writing, and keep the answer. Which channel is rated, what is its Performance Level, and what happens when the model is wrong.

— Jherrod Thomas, The Lion of Functional Safety™