Inverse transformation

Get traj idx via weighted sampling.

Returns a quaternion that generates the given axis-angle rotation. Returns the quaternion: sin(angle/2) * axis + cos(angle/2). Args: axis: A

Returns a random 3D vector with unit length. Generates a 3D vector selected uniformly from the unit sphere. Returns: A normalized 3D vector

:param node_names: a list of names for each tree node :type node_names: List[str] :param parent_indices: an int32-typed tenso

Apply curriculum masking to future motion observations.

Build future motion observations from unified motion data.

Computes the derivative of a sequence of SO3 rotations using central differences. Args: rotations: Quaternion rotations w

Creates environments: 1. loads the robot URDF/MJCF asset, 2. For each environment 2.1 creates the environme

Adds a ground plane to the simulation, sets friction and restitution based on the cfg.

Adds a heightfield terrain to the simulation, sets parameters based on the cfg.

Adds a triangle mesh terrain to the simulation, sets parameters based on the cfg. Very slow when horizontal_scale is small

Store indices of different bodies of the robot

Samples heights of the terrain at required points around each robot. The points are offset by the base's position and rotated by the base

Calculate current motion state masking probability

Sets a vector used to scale the noise added to the observations. [NOTE]: Must be adapted when changing the observations structure

Get unified motion data for both privileged and future frames in a single sampling call. Returns processed motion data that can be used by bot

Initialize torch tensors which will contain simulation states and processed quantities

Returns points at which the height measurments are sampled (in base frame) Returns: [torch.Tensor]: Tensor of shape (num_envs, s

Callback called before computing terminations, rewards, and observations Default behaviour: Compute ang vel command based on target and h

Callback called before computing terminations, rewards, and observations Default behaviour: Compute ang vel command based on target and h

Callback called before computing terminations, rewards, and observations Default behaviour: Compute ang vel command based on target and h

Callback allowing to store/change/randomize the rigid shape properties of each environment. Called During environment creation.

Resets ROOT states position and velocities of selected environmments Sets base position based on the curriculum Selects rando

Update the difficulty of motions for adaptive curriculum learning. This function adjusts the difficulty of each motion based

Implements the game-inspired curriculum. Args: env_ids (List[int]): ids of environments being reset

Linearly interpolate between two frames, including orientation. Args: frame0: First frame to be blended corresponds to (blend = 0

Returns the heading of a rotation q, specified as a quaternion. The heading represents the rotational component of q along the vertical axis (z a

Check if environments need to be reset

Override to include future motion observations while maintaining compatibility.

Computes observations

Basic convolutional encoder for TD-MPC2 with raw image observations. 4 layers of convolution with ReLU activations, followed by a linear layer.

Convert a heightfield array to a triangle mesh represented by vertices and triangles. Optionally, corrects vertical surfaces above the provid

Convert g1_stu_future student policy to TorchScript.