From falling over to factory work: what humanoids can actually do

A decade separates humanoids tumbling at a disaster-response trial from a robot placing sheet-metal parts inside a car plant. That is real progress. It does not make every stage demonstration a factory result, or every factory pilot a proven business case. The useful question is narrower: what was done, under which conditions, and how often?

2015: finishing the course was progress

At the DARPA Robotics Challenge Finals on June 5–6, 2015, robots walked, climbed stairs, turned valves and used tools in tasks modelled on disaster response. DARPA’s pre-event account described human-supervised systems using partial autonomy, with degraded communications built into the test. Its report from the finals recorded both successful runs and conspicuous falls. The achievement was integrating perception, control and mobility under stress. It was not a demonstration of a general-purpose worker acting independently.

2026: scale moved to the arena

On August 15, Beijing’s government reported registrations for the World Humanoid Robot Games: 2,056 robots and 666 registered teams, with 51 events split between 30 competitive events and 21 scenario events. The announcement scheduled the games for August 22–26. Those numbers describe preregistration, not participation. Competition results can show performance on defined tasks; they do not establish that every robot completed them autonomously. The register keeps current product evidence separate for systems such as Unitree G1 and the electric Atlas.

Factory work is narrower, and stronger

BMW reported that its Figure 02 pilot ran during 2025 for about 1,250 operating hours. The robot placed more than 90,000 sheet-metal components before welding and supported production of more than 30,000 BMW X3 vehicles. That is stronger evidence than a stage demonstration because it records repetition in a production setting. It does not mean Figure 02 built those cars independently. Figure’s November 2025 account stated an 84-second cycle time, greater than 99% success and zero interventions as goals; it did not publish them as achieved results. Neither source establishes achieved cost, return on investment, intervention rate or uptime.

What the evidence supports

Movement demonstrations now show capabilities that looked fragile in 2015. Narrow production pilots show that selected tasks can be repeated in real facilities. Proven economics would require achieved operating data and costs that are still not public here. Humanoidrecord’s Shipping Reality view and Figure 03 and Unitree G1 comparison keep those evidence levels visible. Figure 03 is the successor recorded in the register; evidence from the Figure 02 pilot does not carry over to that separate model.

The smallest defensible promise today is therefore specific: a humanoid can perform some bounded tasks, in some environments, with results that depend on the operating setup. Broader claims need broader evidence. That discipline makes progress easier to see without inflating what it means. Until comparable hours, failures, human interventions and costs are published, the distance between an impressive run and dependable work remains measurable but unresolved.