How to record human interventions in a humanoid demo

A humanoid can finish a filmed task while people still prepare parts, approve a restart or resolve an exception outside the frame. For an operations buyer, an intervention log is a way to make that surrounding work visible. It should describe the task, the accepted output and the people needed to keep the workflow moving.

This is an original evidence-accounting guide for an owner, production lead or robotics program manager. It uses dated primary accounts of US warehouse and automotive work. We have not observed those sites, measured their videos or obtained their shift logs. The worksheet is for procurement and review, not instructions for operating a robot or responding to a hazardous stop.

Keep the model, task and period together

Figure's November 19, 2025 account of its Figure 02 BMW program defines interventions as human pauses or resets and describes zero per shift as a goal. It also discusses cycle time and placement accuracy separately. That goal is not a published count of achieved interventions. The account describes sheet-metal handling and says Figure 02 was returning to headquarters as retirement began.

A later model should have its own evidence record. Figure's June 30, 2026 announcement identifies Figure 03 and Helix 02 in a first demonstration of a logistics sequencing workflow at BMW's Spartanburg plant. It describes handling parts and pulling a cart. The source does not publish shift-level intervention or recovery-labor results for that demonstration.

BMW's June 25 customer announcement distinguishes the completed Figure 02 pilot under production conditions from the new Figure 03 project. Its proposed logistics flow also includes onward movement by an automated tugger or Smart Transport Robot and delivery to assembly employees. A task performed by the humanoid is one part of that system.

For your comparison, do not attach an older model's cumulative output to a newer model's demonstration. Preserve the task, hardware, software, site and reporting period in the same row. Mark missing values as undisclosed.

An expansion announcement is a different evidence state

Agility and GXO's October 26, 2026 announcement describes an existing tote-to-conveyor workflow at Flowery Branch, Georgia. It separately says deployments at Statesville, North Carolina and Lebanon, Tennessee are expected in early 2027 using Digit 4, with Digit 5 following when commercially available. Those additional sites are announced plans in that source, not completed deployments.

The release reports cumulative tote movement at the existing site but does not disclose the per-shift intervention and staffing record needed for the worksheet below. A cumulative output claim therefore cannot answer how much assistance your proposed shift might require.

These records support different, bounded conclusions. A customer account corroborates the setting and workflow. A vendor demonstration shows a described task in stated conditions. A future rollout describes intent. None is a substitute for the operating evidence needed at your own site.

Agree what counts before watching the demonstration

Use the log to record human involvement without forcing every event into a single autonomy score. Start with a written task boundary. State where a task begins, what makes its output acceptable, which upstream and downstream steps are excluded and who judges acceptance.

Keep routine setup separate from an unexpected exception. Keep an ordinary material replenishment separate from a person finishing a task the robot could not complete. Preserve the supplier's own definitions alongside yours so a later comparison can explain any difference.

Planned preparation. People, elapsed time and work needed before the observed task starts. This is operating work even if no exception occurs later.
Pause or reset. The observed interruption, who authorized or performed the response, and its recorded scope. The log does not prescribe how to restart equipment.
Remote assistance. Whether a remote person advised, approved, selected an action or directly controlled a step, where disclosed. No local touch does not establish no human involvement.
Manual completion. Work a person performs to finish, correct or replace the attempted task. Accepted output after rework differs from accepted output without rework.
Material or interface exception. The missing input, blocked handoff or connected-system issue reported by the responsible team. Attribute the cause only when evidence supports it.
Maintenance or changeover. Planned and unplanned work, responsible people and time outside the observed task window. A short demonstration does not measure a full shift's support burden.
Unobserved interval. A cut, omitted period or missing record that limits the review. An unknown event count stays unknown.

This classification is our proposed decision aid, not an industry standard or a vendor's measured result. A site can adapt it to its own evidence needs. The responsible specialists retain control of operating procedures and safety acceptance.

The smallest useful entry has a clock and an outcome

For every observed interruption, retain the run identifier, task attempt, model and software revision, event time, observed reason, people involved, local staff time, remote staff time and waiting time. Include the point at which accepted work resumed and whether a person reworked any output. Keep the original observation separate from a later explanation supplied by the vendor.

Elapsed downtime and staff effort answer different questions. If two people attend one interruption, their combined staff time differs from the duration on the clock. If a remote specialist helps while a local worker waits, retain both activities. Do not infer the remote person's time from the visible pause.

Use the same definition of a task attempt throughout the comparison. Record attempts that were abandoned or completed manually, not only successful cycles. Show accepted output separately from attempted output and identify the reviewer. A denominator that excludes difficult attempts can hide the very work the buyer needs to understand.

A hypothetical row, without an invented performance claim

Suppose a demonstration includes a part-handling attempt, a pause and a restart. The log would identify the attempted task and observed pause, name the people whose actions were visible, and leave undisclosed remote involvement marked unknown. If someone later completes the part placement manually, record that as manual completion rather than silently calling the original attempt autonomous.

This is a fictional example of how to classify evidence. It supplies no cycle time, intervention rate or acceptable threshold. It also gives no instructions for investigating a stop or restarting a machine. Those decisions belong to the responsible site team.

Ask for comparable evidence before expanding the scope

Request an agreed reporting window with task attempts, accepted output, excluded conditions and a record of human involvement. Ask how exceptions were collected, who reviewed them and which intervals are missing. Keep the physical task and site conditions consistent with the decision you are considering.

A useful report can show the count of each event category, observed operating time, waiting time and staff effort, with the underlying definitions. Where sufficient data exist, calculate measures only against an explicit denominator. Where the data are missing, the next step is an evidence request, not a performance estimate.

Include the work before and after the humanoid's step: presenting material, maintaining fixtures, replenishing stock, inspecting output and handing it to the next process. Distinguish work transferred to a supplier's support team from work that has disappeared. A purchase proposal should name who owns each part of the process.

What these US examples do not establish

The cited announcements do not provide a November 2026 site inspection, a universal uptime figure, a local service commitment or a staffing forecast for another plant. They do not establish that a model handles arbitrary warehouse or factory work. Dated, task-specific evidence is useful precisely because its boundary can be inspected.

Use the humanoid research guide for the wider technology context. Use this log to ask a narrower question: for the work your team actually needs completed, which human contributions were observed, which were reported and which remain unknown?

Sources & evidence

Official sources reviewed November 5, 2026. Vendor materials establish descriptions, not independently measured performance. Original evaluation criteria are our analysis.

Editorial status. Published November 5, 2026. This is source review and original analysis, not a hands-on test. Read our selection, corrections, and update policy ↗