Humanoid Crosses Line – Jobs On Notice

central eye icon with connected technology symbols over a worker in a hard hat — Photo: Panchenko Vladimir / Shutterstock

A humanoid robot just walked itself off a production line, and the race to give these machines useful jobs is now on.

Story Snapshot

XPeng commissioned a humanoid robot production line; the first IRON unit walked off autonomously.
More than 80 percent of core manufacturing steps on the line are automated, per XPeng.
XPeng targets mass production by year-end 2026 and broader deliveries in 2027, starting in its stores and campuses.
Unitree’s G1 Plus adds dual vision, touch sensing, and a hot-swap battery, with a model trained on millions of examples.

Production Crosses From Demos To A Real Line

XPeng announced that its humanoid robot production lines are now active, and that the first IRON robot “autonomously walked off” the line. The company framed this as a world-first for advanced humanoids exiting a commercial line under their own control. XPeng also said that more than 80 percent of core manufacturing processes on the line are automated, signaling an attempt at scale and repeatability beyond lab builds. These are concrete steps toward making robots you can order, not just watch in a video.

XPeng set an aggressive timeline tied to deployment. Reporting on the launch states the company aims to reach mass production by late 2026, with initial use in XPeng’s own stores and campuses and deliveries in China and abroad in 2027. That staged rollout matters. Companies learn faster when robots face real floors, real customers, and real downtime pressures. Internal deployments also let teams control risk while tuning software, parts supply, and service plans.

Hardware Steps Up: Eyes, Touch, And Power

Unitree, another key player, rolled out its G1 Plus upgrade with a binocular and wide-angle dual vision system, touch sensors built into the head shell, and a removable battery for roughly two hours of operation per pack. The company paired the hardware with a six-billion-parameter control model trained on over five million examples and 2,500 hours of real robot data, covering 64 tasks, according to reporting on the release. The push combines better sensing with broader task learning, the classic one-two in robotics.

These moves fit a wider shift from one-off stunts to field trials. XPeng’s plan to place humanoids in its stores and campuses points to real customer-facing tests. That means aisle layouts, reflective floors, shoppers getting in the way, and staff interruptions. Those frictions are where robots prove they can handle the mess of everyday work. The market reward is large for teams that can keep balance, grip, and judgment intact when the scene changes minute by minute.

The Real Question: How Will Robots Know What To Do?

Hardware lets a robot see and feel; policy and training tell it what to do next. The emerging pattern is blended control: foundation models for broad skills, task fine-tuning for local rules, and fallback modes when the plan goes sideways. Unitree’s training scale suggests a bet on data breadth to cover more scenes and hand-object cases. XPeng’s staged deployments can feed logs back into models and planners, shrinking the gap between a glossy demo and a shift-long job.

Standard-setters are also getting to work. Industrial analysis has stressed a gap between flashy prototypes and dependable, cost-justified performance in the field. That context matters for buyers and for the public. A robot does not earn trust by one great clip; it earns trust by completing thousands of shifts without drama. The constructive use of in-store pilots is to turn unknowns into charts: task completion rates, mean time between failures, and clear triggers for human help.

What Strong Evidence Should Look Like Next

The next proof points should focus on simple questions any floor manager would ask. How long can the robot work before it needs help? How often does a person step in? How quickly does the robot recover from glare, occlusion, or a bumped arm? Which tasks are it allowed to refuse, and what happens then? Vendors can answer with side-by-side trials that compare full autonomy to supervised and teleoperated modes under the same lighting, clutter, and foot traffic.

Goldman Sachs has dramatically raised its 2035 humanoid robotics outlook:

• ~$138B market
• ~6.5M units annually
• Up from ~1.4M previously
The bigger opportunity may not be the humanoid manufacturers themselves.
As Physical AI moves from prototypes toward warehouses,… pic.twitter.com/bCyK5HKpan

— Shay Stock Tracker (@Dragon21Soul) September 20, 2026

XPeng’s line commissioning, the “walked off the line” moment, the 80 percent automation figure, and the 2026–2027 rollout plan are clear markers that manufacturing is getting ready for real volume. Unitree’s G1 Plus and its expanded data and task coverage show that sensing and policy are moving together toward practical use. One brief caveat appears in reporting that commissioning does not equal mass production yet, which keeps the spotlight on execution over the next five quarters.