mani_skill.utils.structs.Pose

class mani_skill.utils.structs.Pose(raw_pose: Tensor)

Wrapper around sapien.Pose that supports managing a batch of Poses and flexible creation of them from a variety of sources (list, numpy array, sapien.Pose). This pose object will also return information with a batch dimension, even if it is just holding a single position and quaternion.

As a result pose.p and pose.q will return shapes (N, 3) and (N, 4) respectively for N poses being stored. pose.raw_pose stores all the pose data as a single 2D array of shape (N, 7).

All sapien.Pose API are re-implemented in batch mode here to support GPU simulation. E.g. pose multiplication and inverse with pose_1.inv() * pose_2, or creating transformation matrices with pose_1.to_transformation_matrix() are suppported they same way they are in sapien.Pose.

Pose Creation

To create a batched pose with a given position p and/or quaternion q, you run:

pose = Pose.create_from_pq(p=p, q=q)

p and q can be a torch tensor, numpy array, and/or list, or None.

If p or q have only 1 value/not batched, then we automatically repeat the value to the batch size of the other given value. For example, if p has a batch dimension of size > 1, and q has a batch dimension of size 1 or is a flat list, then the code automatically repeats the q value to the batch size of p. Likewise in the reverse direction the same repeating occurs.

If p and q have the same batch size, they are stored as so.

If p and q have no batch dimensions, one is automatically added (e.g. p having shape (3,) now becomes (1, 3))

If p is None, it is auto filled with zeros.

If q is None, it is auto filled with the [1, 0, 0, 0] quaternion.

If you have a sapien.Pose, another Pose object, or a raw pose tensor of shape (N, 7) or (7,) called x, you can create this Pose object with:

pose = Pose.create(x)

If you want a sapien.Pose object instead of this batched Pose, you can do pose.sp to get the sapien.Pose version (which is not batched). Note that this is only permitted if this Pose has a batch size of 1.

Pose Indexing

You can index into a Pose object like numpy/torch arrays to get a new Pose object with the indexed data.

For example if pose has a batch size of 4, then pose[0] will be a Pose object with batch size of 1, and pose[1:3] will be a Pose object with batch size of 2.

Methods

__init__(raw_pose)
create(pose[, device])	Creates a Pose object from a given pose, which can be a torch tensor, sapien.Pose, list of sapien.Pose, or Pose
create_from_pq([p, q, device])	Creates a Pose object from a given position p and/or quaternion q
get_p()	Returns self.p, the position
get_q()	Returns self.q, the quaternion
inv()	Returns the inverse of this pose
set_p(p)	Sets the position of this pose
set_q(q)	Sets the quaternion of this pose
to(device)	Move the Pose object to a different device
to_transformation_matrix()	Returns the (N, 4, 4) shaped transformation matrix equivalent to this pose

Attributes

device	Torch Device the Pose object is on
p	The position of this pose
q	The quaternion of this pose
shape	Shape of the Pose object
sp	Returns the equivalent sapien pose.
raw_pose