The 6D rotation representation describes a rotation by the first two columns of its rotation matrix. Unlike Euler angles and quaternions, it is continuous: nearby rotations always give nearby numbers. This makes it a good choice for neural network inputs and outputs. It builds on Rotation Matrices.
Problem With Euler Angles
- Euler angles wrap around, for example roll jumps from to
- a robot gripper pointing straight down sits right at that boundary, so a tiny tilt can change roll by almost
- a network trained on such data must learn a jump that is not in the real motion
- quaternions also have a discontinuity for networks, because and are the same rotation
6D Form
- keep the first two columns, 6 numbers
- the third column is not needed, because
- Zhou et al. (2019) showed that rotation forms with 3 or 4 numbers cannot be continuous, and 5 or 6 numbers can
Back To A Rotation Matrix
A network output is not exactly orthonormal, so rebuild with Gram-Schmidt:
Where Euler Is Still Fine
- small rotation deltas stay near zero, far from any wrap-around, so Euler deltas are fine as actions
- the absolute orientation is the part that needs a continuous form
Use In Robot Learning
- a robot state can be position (3 numbers), the two columns (6 numbers) and the gripper (1 number), 10 in total
- the same function must build the state for training and for inference, so the two cannot disagree on the order of the numbers