A robot demonstration shows a movement. A deployment decision asks whether useful work can happen under stated limits, including when something goes wrong. The distance between those two questions is where a buyer needs evidence.
This is for a business owner or investor evaluating a robotics proposal. You do not need robotics training to complete the exercise below: prepare a focused request for the documents a qualified integrator should explain. This article does not teach you to test or certify a machine.
An announcement with work behind it
On October 1, Agility Robotics and FORT Robotics announced a memorandum of understanding, or MOU, to expand safety collaboration around the planned Digit 5. The announcement describes a pendant, communication systems on the robot, and an interface to safety systems outside it. An MOU records intended collaboration; the announcement is not a complete application assessment. [1]
The work did not start with this announcement. The companies describe a multiyear relationship and an interface they jointly developed. Agility also reports an earlier independent field safety evaluation on a customer production line. Those are relevant reasons to investigate the safety case. They do not supply the full evaluation report or establish what it covers for a proposed Digit 5 installation. [1] [2]
A demonstration compresses reality. You see a successful task. You may not see the operating limits, the recovery procedure, the supervising person, or the faults the application was designed to handle. The useful response is neither automatic trust nor automatic dismissal. Ask for the missing context.
Keep three scopes of evidence separate
First, identify the product generation. In its September 15 announcement, Agility attributes more than 65,000 operating hours to Digit 4. It forecasts Digit 5 early access in the first half of 2027 and general availability by the end of that year. These are manufacturer-reported history and expectations, respectively. Predecessor experience can inform engineering; it cannot become operating history for a generation still described in future terms. [2]
Second, identify the boundary of the assessment. A safety component, the complete robot, and the workplace application are three different scopes. A certificate or test result should say which one it covers, which configuration was assessed, who evaluated it, and what conditions or limits apply. A familiar logo cannot answer those questions. [3] [4]
Third, connect the control to the physical response. A stop command begins a sequence; the moving machine must reach a defined condition appropriate to the task and its surroundings. The evidence must account for people, nearby equipment, and recovery responsibilities. That judgment belongs to qualified engineers and the relevant assessments, not to a viewer of a video. [3] [4]
An August 3 version 1 preprint by Siemens-affiliated researchers makes the scope distinction concrete. It studies a Unitree G1 EDU in one semi-enclosed pick-and-place work cell. The authors separate external safety supervision from the robot-side response and do not claim complete-system certified safety performance. This is evidence about that setup. It is not a test of Digit 5 and cannot certify or condemn a different product. [3]
Turn the question into an evidence request
Imagine a warehouse manager considering a humanoid that carries bins near staff. This is an illustrative situation, not a report of an actual deployment. Before committing to a pilot, the manager can prepare a one-page request in an ordinary document. Start with the proposed task, work area, and places where people could encounter the machine.
Add four rows. In each row, name the requested record, the person responsible for explaining it, and the decision it would inform. Use the following as a starting point, not a complete engineering or legal checklist:
| Requested record | Who should explain it | What it helps decide |
|---|---|---|
| Exact robot generation, configuration, task and operating limits | Supplier and qualified integrator | Whether the evidence concerns the proposed machine and work |
| Application risk assessment, fault response and recovery responsibilities | Qualified integrator | Whether the proposed workplace and human exposure have been assessed |
| Certification or evaluation report, scope, conditions and current applicable requirements | Qualified integrator and evaluator | What a favorable assessment does and does not establish |
| Pilot acceptance criteria and available output, intervention, downtime and cost records | Customer or operator | Whether useful work meets a defined operational and economic test |
OSHA’s technical manual provides a useful reference for application-specific assessment and site acceptance. It contains older standards references, so ask the qualified integrator to identify the current standards and requirements applicable to this installation. The manual is guidance, not a certificate for the proposed humanoid application. [4]
The expected result is a request that connects each claim to a document, a scope, and a responsible reviewer. If a record is unavailable, mark that claim unverified and leave that part of the decision open. Do not improvise a physical stop test or expose staff to the machine to fill the gap.
The commercial thesis still needs a test
Safety integration may create value for suppliers and integrators. That does not establish it as the dominant deployment bottleneck or guarantee attractive returns. Existing safety suppliers may already serve much of the need. Hardware availability, task reliability, commissioning costs, and customer demand may explain more of the economics.
An investor should ask whether repeated deployments reduce commissioning effort while preserving assessed safety and accepted output. If measured delays and costs are dominated by hardware or task capability, weaken the thesis that safety integration is the binding constraint. These sources do not rank those constraints or prove pricing power.
What would change the picture? A sufficiently scoped evaluation report, acceptance evidence from the proposed application, and operating records that make output, interventions, cost, and recovery responsibilities visible. An announced partnership can advance that work. It does not substitute for its receipt.
Intelligence proposes a movement. The people deploying it remain responsible for the consequence.
Primary sources
Evidence reviewed October 5, 2026. Manufacturer statements and forecasts are attributed to their original sources. The cited preprint concerns one setup and is not a certification. This is educational analysis. Qualified professionals must assess the actual machine and application. AI tools assisted research and production; Marc Theiler retains editorial responsibility.
Follow the evidence into practice.
Source-linked As Above analysis on technology, capital and human agency.