Take the safety fence off a humanoid and you don’t get a simpler cell. You inherit the robot safety test plan for the perception models and the independent controller that replaced the cage, and you have to re-prove it on every model and software release. A square of yellow floor tape is not a control.
I keep picturing the cell after the installers leave. Four strips of tape where the posts were. A coffee mug in the middle of the square. Nobody drinks from it. Nobody moves it.
What happens to robot safety when you take the fence away?
On September 15, Agility Robotics unveiled Digit 5 as its first humanoid “engineered for cooperatively safe work at scale,” able to work near people “without the physical safety barriers required by traditional automation.”
That sentence is the whole product. Digit 4, according to co-founder Jonathan Hurst in IEEE Spectrum, had to work behind physical barriers, separated from people at all times, to meet insurance and regulatory demands. He called that a primary design goal to undo. Pras Velagapudi, Agility’s CTO, told Ars Technica that for pallet swaps and shared aisles, “it’s not really practical to constrain the robot within a physical safety cage.”
I believe them about the problem. A cage that a walking robot has to stay inside is a robot that can’t do the job you bought it for. I also believe the next sentence, which Agility does not put in bold: you now own a software safety case.
A fence is a mechanical control. You walk it. You check the interlock. You recertify when someone drills a hole in it. You don’t recertify because a model checkpoint landed on Tuesday.
A perception stack is a different instrument. Agility says Digit 5 uses “proprietary AI algorithms and multiple sensor technologies” to watch for people, then avoid them, stop, or sit down. Visual and audio cues advertise intent. An independent safety controller oversees the response when someone is too close. Hurst has said the sit-down path is conservative on purpose: set the payload down, sit, cut power to the motors before a person can touch the robot.
That’s a state machine glued to a detector. The detector is the part that used to be steel.
Agility also says Digit is the first humanoid to pass an independent field evaluation against industrial safety standards administered by OSHA, and that it is contributing to ANSI/A3 TR R15.108 and ISO 25785-1. Those are real efforts. They are not a substitute for the eval you run on this build, in this cell, with this lighting. Digit 5 is still ahead of us. Early access is aimed at the first half of 2027.
So the claim to take seriously is not “the fence is gone.” It’s “the fence is now a model, and the model will move.”
How do you test a perception model that replaced a cage?
On a test floor, when you swap a fixture you don’t reuse last month’s traveler. You run the sequence that belongs to the new fixture. Same board, new contacts, new limits if the fixture changed the measurement.
A perception update is a fixture swap. Same robot body. Different thing deciding whether a person is in the cell.
The tests that used to be about the cage don’t transfer:
- The cage didn’t have a false negative rate. A missing panel is visible. A missed person in a camera crop is not.
- The cage didn’t care about high-viz versus a gray hoodie. Lighting, clothing, occlusion, and dock-door glare are now in the spec.
- The cage didn’t sit down. Hurst’s sit-and-cut-power path is a behavior. You have to measure time-to-sit, whether the payload is set down first, and whether the motors actually drop.
- The cage didn’t share a computer with a planner. Agility’s architecture, per Hurst, puts safety on an islanded controller with simple verified rules, now aimed at NVIDIA IGX Thor and Halos. That split is the right idea. You still have to prove the island still trips when the main stack is wedged, not only when the demo is clean.
This is the same discipline I wanted in who signs the test report for a robot: pass criteria set before the run, a method another person can reproduce, failures kept in the report. Simulation red-teams are useful. They are not a walk-around of the tape square.
It is also the same lesson as your tests are lying to your agent. If the detector’s eval set doesn’t include the way people actually cut through that aisle on night shift, the eval is a flaky suite with the polarity flipped. It stays green while the instrument is blind.
What I’d actually measure, before anyone talks about cooperative safety as a feature: false negatives first, detect-to-act time against the old cage’s stop, the islanded controller with the main model killed, and regression on the last signed clip set. One run is an anecdote. A drop that improves tote-pick and misses a crouched person is a failed release. If the safety computer needs the perception model to be healthy, you don’t have two systems. You have one system with a backup logo.
What belongs on a robot safety test plan after the fence comes down?
Here is the traveler I’d clip to the cell. The numbers are an example. The rows are not. Fill them in for your site before the tape comes up.
Release gate: humanoid, perception safety stack
Build: perception 2026.11.17 / controller 5.2.1 / skills 5.2.1
Cell: 4. Yellow tape square on the old cage footprint.
Pass criteria (frozen 2026-11-10, before this run)
[ ] False negative on labeled site video: 0 on the hold-out set
[ ] Detect-to-sit, person at 2 m walking in: at most 400 ms to motion change
[ ] Independent controller trips with main model process killed
[ ] Night-shift lighting pack: dock glare, overhead out, high-viz on/off
[ ] Occlusion pack: person steps out from behind a tote stack, 20/20
[ ] Clothing pack: vest, hoodie, hairnet, empty-handed, pushing a cart
[ ] Payload down before sit, on 10/10 loaded approaches
[ ] No regression vs last signed build on the frozen clip set
[ ] Audit log names model hash, controller version, station, operator
Results
hold-out FN: ___ / ___
detect-to-sit p95: ___ ms
islanded trip: pass / fail
clip-set regress: ___ worse, ___ better, ___ same
signed: [name, independent of model team]
date:
Do not ship if any box is blank.
Coffee mug remains in the square. Calibration witness.
That last line is a joke until it isn’t. A physical witness object in the old cage footprint is a cheap check that the cell geometry you certified is the cell you still have. If someone drags a pallet into the square, the tape lied.
Notice what is not on that traveler: “the model feels safer.” “Demo passed with a visitor in the aisle.” “OSHA evaluation from a previous generation.” Those can be inputs. They are not this release.
For a model drop, the clock starts at the last signed build and ends when this build has passed the frozen suite in the real cell, not when the weights finished copying. Same idea as changeover time: last good part to first good part at rate.
Uptime dashboards still matter. I would put a humanoid on the same SLOs I’d use for a service. They do not replace the safety traveler. Uptime of an unsafe detector is a longer exposure. Digit 5’s 20-hour duty cycle makes that worse, not better: more hours next to people is more hours the detector has to stay honest.
Why do you have to re-prove safety on every model release?
Because the control you inspect changed from a weldment to a checkpoint.
On a cage, the recert trigger is physical: damage, a bypass, a layout change. On a perception stack, the recert trigger is any change that can alter detection or the sit/stop path. That includes:
- A new perception model, even a “small” fine-tune on site data
- A prompt, policy, or planner change that can keep the arms up longer
- A controller firmware drop
- A camera, lens, or mount swap
- A lighting change in the cell
- A new skill that carries a payload through the old cage footprint
Agility’s own split tries to contain this. Hurst has said a dropped network connection should not take away on-robot safety, and that the safety computer has relatively simple, verified rules. Good. Keep it that way. Then treat every change outside that island as a change to the thing the island is watching.
The industry temptation will be to ship skills faster than safety evals. That’s how you get a tote-pick improvement and a crouched-person miss in the same tag. Gate the skill the way you’d gate a torque-limit change. If the perception hash moved, the traveler runs. If the traveler isn’t signed, the robot stays in the square.
Keep a person on the first shifts after a drop, the same way a new test program gets a babysitter. When the person and the detector disagree, the detector is wrong until proven otherwise. Log both.
Until a standard form exists, the plant that runs the robot owns the gap. Insurers already forced Digit 4 behind a fence. They will not take “the model is better now” as a substitute for a named suite.
The tape is still on the floor so you remember what used to do this job.
Related
- Agility, Digit 5 announcement, Sep 15, 2026
- Jonathan Hurst, Digit 5 redesign, IEEE Spectrum
- Ars Technica, Digit 5 stop and squat
- Who signs the test report for a robot?
- Monitor a humanoid like a production service