Humanoid robots

A humanoid robot is a machine built roughly to the proportions of a person: two legs or a wheeled base, two arms, hands or grippers, a head carrying cameras, and a centre of mass it has to keep over its feet. The definition sounds like a design choice. It is really an economic argument.

Why the human shape at all

Wheels are more efficient than legs. A robot arm bolted to a bench is cheaper, faster and far more reliable than an arm attached to a swaying torso. Nobody builds a humanoid because the shape is good engineering.

They build it because the world is already built. Doorways, stairs, shelf heights, tool handles, vehicle cabs, light switches and packing boxes were all sized for a body about 1.7 metres tall with two hands at the end of jointed arms. A machine with that body can be dropped into a workplace without redesigning the workplace. Every other robot form asks the building to change first.

The counter-argument, made by people who have automated factories for thirty years, is that changing the building is usually cheaper than building a humanoid. That is why warehouse automation went to conveyors and shuttles rather than androids, and it stays true wherever the task is repetitive and the layout is yours to control. The humanoid case is strongest exactly where it is not: an existing building, a changing task, and a job description written for a person.

The machines that exist

The field, by how far it has actually got

Atlas from Boston Dynamics is the all-electric successor to the hydraulic machine from the viral videos. Boston Dynamics describes it as a commercial humanoid for material handling that navigates to a charging station and swaps its own battery, with skills that transfer across a fleet, and says it is in field testing at a Hyundai facility on sequencing tasks.

Digit from Agility Robotics is the one the company calls commercially deployed, run by a cloud platform it calls Arc, with Schaeffler, GXO and Toyota Motor Manufacturing Canada among the names on its own customer wall.

Figure 03 is described by Figure as a general-purpose humanoid for every day, driven by its Helix model, and demonstrated in June 2026 performing a logistics sequencing workflow in Hall 52 at BMW Group Plant Spartanburg.

Apollo 2 from Apptronik comes in bipedal and wheeled configurations, with swappable batteries the company says enable 7x22 operation, and it is the machine Google DeepMind used to demonstrate whole-body control in July 2026.

Optimus is Tesla’s, described on Tesla’s own AI page as a goal rather than a product: a general-purpose, bipedal, autonomous humanoid for unsafe, repetitive or boring tasks.

G1 from Unitree is the one you can buy. Unitree markets it as a humanoid agent with 23 to 43 joint motors depending on configuration, an optional force-control dexterous hand, and a price from US$13,500, which is roughly two orders of magnitude below where humanoids sat five years ago.

NEO from 1X is aimed at the house rather than the factory, sold on a $200 deposit under an early-access programme, with the company saying it is built for full autonomy and that capabilities unlock as it learns.

What they can do, and where the demonstrations stop

Walking is solved well enough to be boring. Balance, stairs, uneven ground, recovering from a shove: all of that works, and has for several years.

Manipulation is where the field actually is. Picking up a rigid box with a two-fingered gripper is reliable. Tying a knot, sealing a bag, threading a cable or handling something floppy with five fingers is not. When Google DeepMind published its whole-body control results in July 2026, its own chart put whole-body and gripper tasks at a medium to high success rate and said multi-finger dexterous manipulation remained challenging. That single sentence is the most accurate summary of humanoid capability available.

Speed is the other quiet limit. A humanoid that completes a task correctly at a third of human pace is a research success and a commercial problem, because the business case is priced against a wage.

What has to happen next

Three things stand between the current pilots and anything resembling a workforce.

Hands have to become both dexterous and durable, which are opposing requirements: more joints means more to break. Models have to reach fine manipulation, which is where robot foundation models are aimed. And the safety standards have to arrive, because a legged machine that can fall over is a hazard class that existing industrial robot standards were not written for, which is what the safety standards work is now dealing with.

None of those is a research mystery. All of them are years of engineering, and the field is currently spending its credibility ahead of them.