Rotation sense and the unwind back to rest are both in the definition
Aliases: wrist roll · pronation and supination · palm flip · rotational return
What it is
A palm flip, wrist roll, or turning of the hand must state two things in its definition: which way it turns, and how it gets back to a pose from which it can turn again. “Give it a turn” is not a definition. Pronation and supination are opposite angular velocities to the sensor; if the unwind is not inside the same token, the reverse rotation ships as a second command. A swipe's round trip lives in translation. This one lives in rotation about the long axis of the forearm. Both have to close; they close in different degrees of freedom.
Why it happens
Forearm pronation/supination has about 180° of range; an everyday “flip the hand” often uses only a slice. The sign of angular velocity is the direction. Skeleton tracking that only looks at start and end palm facing drops the sign along the way, and sees a round trip as “palm up, then palm up again.” The return is almost necessarily a reverse rotation: to restore the starting facing, the angle just travelled has to be walked back. If the recognizer already committed on outbound angular velocity, inbound reverse velocity is the mirror token. Unlike pinch, rotation has no fingertip closure as an edge; edges come from angular velocity crossing zero or facing returning to rest. Gravity pulls the forearm toward a cheaper supinated angle, so the return is not fully voluntary. Templates must count that “gravity hauled it back” stretch as part of the definition, or everyone's return looks like noise.
Studying it
Label the sign of angular velocity on outbound and return, and see whether events commit outbound, on return, or twice. Factors: amplitude, speed, whether return to initial facing is required. Measures: direction accuracy, return false fires, whether the palm ends at the rest facing. An inertial unit on the forearm is a better gold standard than vision alone for separating true rotation from apparent rotation caused by the camera orbiting the hand. Do not control with two-finger rotate on a touchscreen: that is a two-point path in a plane, not a forearm axis.
Where it stops holding
When the object itself is turning (unscrewing a cap, driving a virtual knob), the hand need not return to the initial facing; the unwind is not part of this token, and the mapping should be continuous angle. Headsets looking at the hand see forearm rotation. A room camera from the front sees little silhouette change in pronation/supination, so direction smears. Shoulder abduction plus a flip injects swipe-like translation; the classifier must split rotation from translation or one action lights both families. People with limited range cannot fill the defined angle; allow a smaller amplitude rather than fail them.
Applying it
- Write axis, positive sense, whether the unwind belongs to the same token, and the commit point, for every rotation gesture.
- Mirror vocabularies must be regressed on full outbound-plus-return, counting reverse events.
- Teaching runs until rest facing is restored. If the function is a continuous twist, say “no reset; release to stop.”