Humanoid robots: 7 tests that separate a demo from real work

A robot video can be authentic and still create a misleading impression. Editing can hide failures, the environment can be prepared down to the millimeter, and an unseen operator can intervene between shots. Yet dismissing every demonstration as marketing also obscures genuine engineering progress.

On June 30, 2026, Figure showed Figure 03 inside Hall 52 at BMW's Spartanburg plant. The robot sequences parts, repositions its body and pulls a wheeled cart. The publication is useful because it can be tested against a simple framework: what do we know about the location, task, autonomy, repetition and cost?

1. A named location and customer

A scene filmed in an anonymous studio mostly proves that a prototype exists. A named industrial site adds accountability: the customer, workflow and equipment can be checked against the story. Figure identifies BMW Spartanburg and a specific hall. That is a credibility signal, though not an independent audit.

2. An end-to-end task

“Working in a factory” is meaningless without a unit of work. Figure describes a sequencing workflow: picking components, ordering them, moving the body and pulling a cart. A defined task makes it possible to ask about throughput, errors, recovery and accepted quality.

The workflow also matters because a conventional fixed arm may struggle when the task spans shelves, carts and workstations designed around people. Humanoid form is only an advantage if it avoids expensive changes to the environment.

3. The level of human assistance

The decisive question is rarely whether the robot moves on its own. Ask who starts the mission, who confirms the next step, who corrects an error and whether remote assistance is available. Figure presents Helix 02 as full-body autonomous control, but it does not publish interventions per hour for this BMW case.

An honest deployment report should separate setup, supervision, remote guidance and physical recovery. Those roles have very different costs.

4. Continuity, not the best moment

A continuous sequence carries more evidence than a compilation of isolated gestures. Helix 02 is designed to coordinate walking, balance and manipulation without rigid hand-offs. But a successful multi-minute run does not show that the system can sustain a full shift.

The next useful evidence should cover accumulated operating hours, different shifts, varied loads and changes in the surrounding workflow.

5. Success rate and recovery

A useful system must do more than succeed. It must detect failure and return to a safe state. Ask for the percentage of accepted cycles, failure categories, mean time between interventions and average recovery time.

Without those values, buyers cannot compare a humanoid with a fixed robot, an autonomous mobile robot or a human-operated process. A spectacular hard task may still deliver less value than a modest task completed reliably.

6. Manufacturing the robot itself

In April, Figure said it had delivered more than 350 Figure 03 units and demonstrated a one-robot-per-hour assembly cycle. Moving from a handful of prototypes to a fleet matters because manufacturing variation exposes weaknesses and produces more operational data.

These numbers are company statements. They do not establish broad commercial availability, field reliability or an audited production rate.

7. Cost per useful outcome

Unit price is not enough. A serious calculation adds integration, tooling, safety work, supervision, energy, maintenance, updates and downtime, then divides the total by accepted units of work. Figure does not publish those BMW economics.

Any precise return-on-investment claim would therefore be premature. The buyer needs a measured pilot with written success and exit criteria.

How to read the Figure 03 evidence

The BMW case matters more than a viral dance because it combines a named industrial customer, a material workflow and whole-body coordination. What remains missing is equally clear: success rate, accumulated operating time, interventions, incidents and cost per operation.

The verdict is neither “the revolution is complete” nor “it is only a trick.” Figure 03 has crossed a verifiable threshold while generalization remains unproven. That distinction is how readers can follow robotics without hype or reflexive cynicism.