Transform.inv seems incorrect

[simon-bachhuber]

Dear Brax authors,

while scrolling through some of the source i stumpled upon

class Transform(_Base):
  def inv(self):
    """Invert the transform."""
    return Transform(pos=-1.0 * self.pos, rot=math.quat_inv(self.rot))

Is this really correct? Is it due to some convention i am unaware of?

It also seems inconsistent with Roy Featherstone

Greetings

[btaba]

Hi @SimiPixel we follow a similar convention to TDS, where Transform is expressed as a translation and quaternion rotation, rather than a 6x6

[btaba]

Hi @SimiPixel , indeed thanks so much for pointing this out!! inv is clearly wrong, it should be:

rot = math.quat_inv(self.rot)
Transform(pos=math.rotate(-1.0 * self.pos, rot), rot=rot)

as in https://github.com/erwincoumans/tiny-differentiable-simulator/blob/master/src/math/transform.hpp#L191-L195

where we are also using right associative transforms (hence the slight diff with the screenshot you provided)

[simon-bachhuber]

No worries. As long as

t = Transform()
t.inv().do(t) == t.do(t.inv()) == Transform.zero()

i am convinced 👍