Humanoid Robots in 2027: What Will They Actually Be Able to Do?

01 / The story

What is actually working today

The clearest evidence comes from logistics. GXO described deploying Agility's Digit in a live operating facility in August 2024, and Agility reported in November 2025 that Digit had moved more than 100,000 totes at GXO's Flowery Branch site. In September 2026, Agility announced the Digit 5, citing over 65,000 hours of accumulated operational experience and more than $300 million in multi-year orders. Those hours are vendor-reported and span the fleet's history, not one robot—and orders are not completed deliveries. [1, 2, 3]

In automotive, Figure reported in November 2025 that its Figure 02 robots had loaded over 90,000 parts across 1,250+ runtime hours at BMW, contributing to the production of more than 30,000 X3 vehicles, with 10-hour weekday shifts described. The same announcement said Figure 02 was returning from BMW as that generation retired—so historic results should not be conflated with current-generation deployment. This was parts loading, not autonomous manufacture of entire cars. [4]

Apptronik's Apollo robots have been present at Mercedes-Benz sites in Berlin and Kecskemét since March 2025, and Apptronik's June 2026 announcement describes Apollo 2 and its Robot Park as platforms for data collection and training, including fleets at partner sites. Presence at a customer site can mean training and pilot activity rather than autonomous production work. [8, 9, 10]

02 / The story

The Tesla question

Tesla's Optimus draws the most attention and the least verifiable data. A September 25, 2026 Electrek report, citing The Information, described builds of several hundred robots per week in August—mostly for internal testing and training—with tightly controlled, supervised factory tasks, fragile hands and generalization problems. [5]

That is attributed secondary reporting, not an audited production disclosure or independently measured customer operation. A reported target above 1,000 robots per week is not achieved output, and robots built are not robots doing useful paid work. Treat Optimus scale claims as unverified until Tesla or its customers confirm deployments with task counts.

03 / The story

Demos versus deployments

A robot demo proves a robot did something once, under conditions you cannot see. A deployment proves something harder: that a customer keeps paying for the work. The gap between the two is where most humanoid news lives.

Boston Dynamics' September 21, 2026 announcement opened Hyundai's Robotics Metaplant Application Center in Georgia—explicitly a training and testing facility. Hyundai's own September 30 account places initial parts-sequencing work in 2028 and parts assembly from 2030. Announced ambitions for 25,000 units across group plants and a 30,000-per-year factory are future plans, not installed robots. [11, 12]

Unitree's G1 is advertised from $13,500, and its official shop warns that the standard product does not support secondary development—directing customization buyers to the EDU version. Hardware availability and athletic ability do not verify a sustained factory job or an all-in industrial integration cost. [6, 7]

04 / The story

The household reality check

Home robots are the dream and the hardest problem. On September 17, 2026, Figure reported its Helix 2.5 system performing three household behaviors—tidying, towel folding and bed making—tested in 30 previously unseen homes. A detailed account of the company-run evaluation counted 237 successful trials out of 420, about 56%. [14, 15]

That is real progress in generalization: the system transferred to homes and objects it had never seen. It is also plainly not dependable household service—a coin-flip-plus success rate in a company-run evaluation is a long way from a robot you'd trust with your grandmother's kitchen. Meaningful home progress in 2027 is plausible; dependable general household service is not established.

05 / The story

The technical limits that matter

Dexterity: a hand's joint count is not a measure of task success. Figure's published placement and zero-intervention figures at BMW were described as targets, not proof of all-shift achieved performance, and household manipulation proved much less reliable in the later evaluation. [4, 14, 15]

Battery life: separate continuous work per charge from daily productive availability. Digit 5 advertises a 10:1 run-to-charge ratio and 20 working hours per day—vendor specifications, not independently measured customer uptime. Cross-vendor battery figures are not comparable without payload, walking distance and task intensity. [13]

Reliability: cumulative hours and totes moved show repeated execution, but without scheduled hours, intervention logs, maintenance records and fleet size they cannot tell you availability or mean time between failures. [2, 4]

Cost: a $13,500 starting price is a hardware entry point. Integration, safety engineering, supervision, charging, spares and service are priced separately. No independently verified all-in cost per productive hour exists across these vendors. [6, 7]

06 / The story

What 2027 will probably look like

Based on the evidence rather than the announcements, here is a calibrated outlook.

Relatively high confidence: more paid, bounded tote and box transfers and parts loading in structured facilities—Digit and Figure already do this repeatedly. More mobile-manipulation pilots and training programs—Apollo 2 and Atlas programs explicitly emphasize training. [1, 2, 4, 10, 11, 12]

Medium confidence: better battery and charging orchestration in selected settings, and broader home-task transfer with continued supervision needs. The tests to watch: independently measured productive hours at customer sites, and unedited repeated home trials with intervention counts. [3, 13, 14, 15]

Low confidence, not supported: universal unsupervised household workers, or drop-in replacements for entire factory jobs in 2027. The collected evidence establishes narrow tasks and training programs, not dependable general labor. If a headline promises otherwise, ask it for a named customer, a task count and an intervention log.