Glossary
Labels only help if they mean the same thing everywhere. These are the definitions behind every confidence grade, stage and part name on the site.
Confidence labels
- Official
From the company itself: its official website, announcements or data sheet. The strongest grade, but still a company claim.
- Reported
From reliable media reports or measurements taken at a demonstration, without confirmation from the company.
- Rumor
From anonymous sources, social media or speculative articles. Recorded so it can be tracked, never mixed with facts.
- Estimate
Calculated or estimated by Wingchiki, for example by comparing videos. Shown only when nothing better exists.
Development stages
- Development stage
① Concept → ② Prototype → ③ Pilot → ④ Production → ⑤ On sale. Discontinued robots sit outside the five steps. A stage is Wingchiki's reading of public evidence, not a shipment count.
- ① Concept
Only renders or announcements exist; no working robot has been shown.
- ② Prototype
A working prototype has been shown, but it has not gone into customer sites.
- ③ Pilot
Being trialled at real customer sites such as factories or warehouses.
- ④ Production
Volume production has started on a line, or the robot is being delivered in volume.
- ⑤ On sale
A price is public and ordinary customers can order it.
Forms
- Biped
A humanoid that walks on two legs. It can use stairs and thresholds built for people, but balance control is hard and costs are higher.
- Wheeled humanoid
An upper body with two arms and a head on a wheeled base instead of legs. Stable on flat factory and warehouse floors, so it tends to reach the field sooner than bipeds.
Spec terms
- Degrees of freedom (DoF)
The number of joint axes that can move independently. More DoF allows more varied motion but makes control harder.
- Payload
The maximum weight the robot can lift or carry. Makers differ on whether it is per arm or for both arms.
- Runtime
How long the robot can work on one battery charge. It varies a lot with how hard the work is.
Parts
- Actuator
The drive unit that moves a joint. Usually a joint module that combines motor, reducer, sensors and driver, and one of the largest shares of a humanoid's cost.
- Rotary actuator
An actuator that turns a joint around an axis, used in shoulders, elbows and hips.
- Linear actuator
An actuator that turns motor rotation into pushing and pulling motion, used where large force is needed, such as knees and ankles. A roller screw or ball screw sits inside.
- Reducer
Slows a motor's fast rotation while raising torque. Humanoids mostly use harmonic and planetary reducers.
- Harmonic reducer
A reducer in which a thin metal gear flexes as it meshes. Small, light and with little backlash, so it is common in arm and wrist joints.
- Planetary reducer
A reducer in which small gears orbit a central gear. Shock-resistant and relatively cheap, so it is common in leg joints.
- Roller screw
Rollers around a threaded shaft turn rotation into linear motion. It carries more force than a ball screw, so it is used in linear actuators. The planetary roller screw is the main type.
- Quasi-direct drive (QDD)
An actuator design with a low gear ratio. The motor feels outside forces directly, so it handles impacts well and moves smoothly. Common in legged robots such as Unitree's.
- Robot hand
The part that grasps and handles objects. The number of fingers, degrees of freedom per hand and tactile sensing decide which tasks it can do.
- Gripper
A simple hand that holds objects with two or three jaws. Less versatile than a human-like hand, but sturdier and cheaper.
- Tactile sensor
Measures force and pressure on fingers and palms. Used to hold fragile objects and notice slipping.
- Lidar
A sensor that measures distance with lasers, used to map the surroundings and avoid obstacles.
- Depth camera
A camera that also measures the distance to each point, used to understand an object's position and shape in 3D.
- BOM
Bill of materials: the list and cost of parts needed to build one robot. It comes up often in price forecasts.
Physical AI
- Physical AI
AI that sees, decides and moves a body in the real world rather than on a screen. Used to refer to a robot's VLM, VLA and whole-body control together.
- Understanding & planning (VLM)
Vision-Language Model. AI that understands images and language together to read a scene and plan what to do.
- Action generation (VLA)
Vision-Language-Action. AI that takes images and instructions and turns them into arm and hand motions.
- World model
AI that imagines the next scene as video. Used to generate virtual training video or to evaluate the result of an action in advance.
- Whole-body control (RL)
A model that carries out motion goals from higher-level models across all joints without falling. Often learned with reinforcement learning (RL) in simulation; model predictive control (MPC) is also used.
- Teleoperation
A person moves the robot directly with VR gear or controllers. Used for demos, and the recorded motion becomes training data. When watching a video, check whether the motion is autonomous or teleoperated.
- Imitation learning
Learning by copying human demonstrations, mainly used to teach hand tasks from teleoperation data.
- Sim-to-real transfer
Moving motions learned in simulation onto a real robot. The challenge is reducing the drop in performance caused by physical differences between the two.