Humanoid robots have become the most photographed and least understood category in robotics. Almost everything the public sees is a short clip, and a clip is a poor instrument for judging a physical system. If you are trying to track this field seriously — as an engineer, a buyer, or just someone tired of the noise — a small set of questions does most of the work.
1. Which market is this machine in?
There are at least three, and they barely overlap.
Research platforms are experimental substrates. The buyer will replace the controller, bolt on sensors, and deliberately run the thing into failure. What matters is open interfaces, documented kinematics, a simulation model that matches the hardware, and a vendor who tolerates modification. Gait quality matters less than whether you are allowed to rewrite it.
Industrial and logistics platforms are evaluated on cost per unit of work. Duty cycle beats peak payload. Mean time between failures and mean time to repair beat demonstration footage. Battery swap logistics beat headline runtime. And the humanoid form has to justify itself against wheeled bases and fixed arms that are cheaper, older and already certified — the honest argument for legs is that the environment was built for people and cannot be rebuilt.
Consumer and home platforms face the strictest constraints and have the fewest shipping products. A factory cell can be fenced; a living room cannot.
A ranking that mixes these three tells you nothing, because each buyer would order the same list differently and each would be right.
2. Where do the degrees of freedom live?
Total DoF is the number vendors quote first because it is easy to inflate — count each finger joint separately and it doubles. What matters is distribution: legs, torso, arms, hands, neck. A rich hand on a rigid torso solves different problems than a mobile waist with simple grippers.
3. Does every number come with conditions?
- Payload without arm extension and duty cycle is a peak, not a working rating.
- Runtime without an activity is meaningless; idle standing and loaded walking differ by a large factor.
- Autonomy without a task scope, an environment and a success rate over N trials is a word, not a claim.
4. Which autonomy mode produced the video?
Teleoperation, shared autonomy and full task autonomy look identical on camera. A robot piloted by a person in a motion-capture rig moves fluidly because a human is producing the motion. If a demo does not state the mode, assume the one most favourable to the manufacturer.
5. How does it fail?
Failure mode matters more than success rate. A machine that fails by stopping is deployable. A machine that fails by falling into a shelf is not — at the same percentage.
6. Is there a service story?
Feet, harnesses and actuators are consumables on a walking machine. Published service intervals, a spare-parts ordering path and field-replaceable modules decide whether a pilot becomes a deployment. This is the most under-reported dimension in humanoid coverage.
7. Can it actually be bought?
Announcement, prototype, pilot and general availability are four different states that press coverage flattens into one. Before treating a price as real: have units shipped to customers outside the manufacturer, is there a quote you could act on, and what is the lead time?
None of this argues the category is fake. It argues that it is early, and that early markets reward careful readers. Applying these seven questions to any two platforms usually turns a marketing draw into a clear difference.
Model-by-model breakdowns built on this framework are maintained at Humanoza.
Top comments (0)