toolbox: publish actions.py

toolbox/actions.py

@@ -1,7 +1,14 @@
 """Derive action targets from the handheld sensor poses.
 
 React is a handheld dataset — there is no robot command. Actions are derived
-from the OptiTrack 6-DoF sensor poses (xyz + quat wxyz) stored per frame.
+from the OptiTrack 6-DoF sensor poses stored per frame as
+[x, y, z, qx, qy, qz, qw] — position in METRES, quaternion SCALAR-LAST, the
+`scipy.spatial.transform.Rotation.from_quat` convention. An earlier version of
+this docstring said "wxyz"; the code was always scalar-last, so the text was
+the error, but a reader who trusted it would have swapped w into x.
+
+Rotation deltas here are WORLD-FRAME: dq = q[i+1] * q[i]^-1, integrated as
+q[i+1] = dq * q[i]. Round-trip verified to 6e-9 m and 1.4e-5 deg.
 """
 from __future__ import annotations
 
@@ -22,8 +29,8 @@ def delta_pose_action(poses):
     """Frame-to-frame delta: translation diff + relative rotation (quat).
 
     Returns (T, 7): [dx,dy,dz, dqx,dqy,dqz,dqw], last row zero-translation +
-    identity rotation. Quaternions assumed (w,x,y,z)? -> stored as xyz+quat;
-    here treated as [x,y,z, qx,qy,qz,qw] (scalar-last), matching schema docs.
+    identity rotation. Quaternions are SCALAR-LAST throughout, and the rotation
+    delta is world-frame: dq = q[i+1] * q[i]^-1.
     """
     p = np.asarray(poses, np.float64)
     T = p.shape[0]