jaxsim 0.2.dev191__py3-none-any.whl → 0.2.dev366__py3-none-any.whl

jaxsim/rbda/forward_kinematics.py

@@ -0,0 +1,113 @@
+import jax
+import jax.numpy as jnp
+import jaxlie
+
+import jaxsim.api as js
+import jaxsim.typing as jtp
+from jaxsim.math import Adjoint, Quaternion
+
+from . import utils
+
+
+def forward_kinematics_model(
+    model: js.model.JaxSimModel,
+    *,
+    base_position: jtp.VectorLike,
+    base_quaternion: jtp.VectorLike,
+    joint_positions: jtp.VectorLike,
+) -> jtp.Array:
+    """
+    Compute the forward kinematics.
+
+    Args:
+        model: The model to consider.
+        base_position: The position of the base link.
+        base_quaternion: The quaternion of the base link.
+        joint_positions: The positions of the joints.
+
+    Returns:
+        A 3D array containing the SE(3) transforms of all links belonging to the model.
+    """
+
+    W_p_B, W_Q_B, s, _, _, _, _, _, _, _ = utils.process_inputs(
+        model=model,
+        base_position=base_position,
+        base_quaternion=base_quaternion,
+        joint_positions=joint_positions,
+    )
+
+    # Get the parent array λ(i).
+    # Note: λ(0) must not be used, it's initialized to -1.
+    λ = model.kin_dyn_parameters.parent_array
+
+    # Compute the base transform.
+    W_H_B = jaxlie.SE3.from_rotation_and_translation(
+        rotation=jaxlie.SO3.from_quaternion_xyzw(xyzw=Quaternion.to_xyzw(wxyz=W_Q_B)),
+        translation=W_p_B,
+    )
+
+    # Compute the parent-to-child adjoints and the motion subspaces of the joints.
+    # These transforms define the relative kinematics of the entire model, including
+    # the base transform for both floating-base and fixed-base models.
+    i_X_λi, S = model.kin_dyn_parameters.joint_transforms_and_motion_subspaces(
+        joint_positions=s, base_transform=W_H_B.as_matrix()
+    )
+
+    # Allocate the buffer of transforms world -> link and initialize the base pose.
+    W_X_i = jnp.zeros(shape=(model.number_of_links(), 6, 6))
+    W_X_i = W_X_i.at[0].set(Adjoint.inverse(i_X_λi[0]))
+
+    # ========================
+    # Propagate the kinematics
+    # ========================
+
+    PropagateKinematicsCarry = tuple[jtp.MatrixJax]
+    propagate_kinematics_carry: PropagateKinematicsCarry = (W_X_i,)
+
+    def propagate_kinematics(
+        carry: PropagateKinematicsCarry, i: jtp.Int
+    ) -> tuple[PropagateKinematicsCarry, None]:
+
+        (W_X_i,) = carry
+
+        W_X_i_i = W_X_i[λ[i]] @ Adjoint.inverse(i_X_λi[i])
+        W_X_i = W_X_i.at[i].set(W_X_i_i)
+
+        return (W_X_i,), None
+
+    (W_X_i,), _ = jax.lax.scan(
+        f=propagate_kinematics,
+        init=propagate_kinematics_carry,
+        xs=jnp.arange(start=1, stop=model.number_of_links()),
+    )
+
+    return jax.vmap(Adjoint.to_transform)(W_X_i)
+
+
+def forward_kinematics(
+    model: js.model.JaxSimModel,
+    link_index: jtp.Int,
+    base_position: jtp.VectorLike,
+    base_quaternion: jtp.VectorLike,
+    joint_positions: jtp.VectorLike,
+) -> jtp.Matrix:
+    """
+    Compute the forward kinematics of a specific link.
+
+    Args:
+        model: The model to consider.
+        link_index: The index of the link to consider.
+        base_position: The position of the base link.
+        base_quaternion: The quaternion of the base link.
+        joint_positions: The positions of the joints.
+
+    Returns:
+        The SE(3) transform of the link.
+    """
+
+    return forward_kinematics_model(
+        model=model,
+        base_position=base_position,
+        base_quaternion=base_quaternion,
+        joint_positions=joint_positions,
+    )[link_index]

jaxsim/rbda/jacobian.py

@@ -0,0 +1,201 @@
+import jax
+import jax.numpy as jnp
+import numpy as np
+
+import jaxsim.api as js
+import jaxsim.typing as jtp
+from jaxsim.math import Adjoint
+
+from . import utils
+
+
+def jacobian(
+    model: js.model.JaxSimModel,
+    *,
+    link_index: jtp.Int,
+    joint_positions: jtp.VectorLike,
+) -> jtp.Matrix:
+    """
+    Compute the free-floating Jacobian of a link.
+
+    Args:
+        model: The model to consider.
+        link_index: The index of the link for which to compute the Jacobian matrix.
+        joint_positions: The positions of the joints.
+
+    Returns:
+        The free-floating left-trivialized Jacobian of the link :math:`{}^L J_{W,L/B}`.
+    """
+
+    _, _, s, _, _, _, _, _, _, _ = utils.process_inputs(
+        model=model, joint_positions=joint_positions
+    )
+
+    # Get the parent array λ(i).
+    # Note: λ(0) must not be used, it's initialized to -1.
+    λ = model.kin_dyn_parameters.parent_array
+
+    # Compute the parent-to-child adjoints and the motion subspaces of the joints.
+    # These transforms define the relative kinematics of the entire model, including
+    # the base transform for both floating-base and fixed-base models.
+    i_X_λi, S = model.kin_dyn_parameters.joint_transforms_and_motion_subspaces(
+        joint_positions=s, base_transform=jnp.eye(4)
+    )
+
+    # Allocate the buffer of transforms link -> base.
+    i_X_0 = jnp.zeros(shape=(model.number_of_links(), 6, 6))
+    i_X_0 = i_X_0.at[0].set(jnp.eye(6))
+
+    # ====================
+    # Propagate kinematics
+    # ====================
+
+    PropagateKinematicsCarry = tuple[jtp.MatrixJax]
+    propagate_kinematics_carry: PropagateKinematicsCarry = (i_X_0,)
+
+    def propagate_kinematics(
+        carry: PropagateKinematicsCarry, i: jtp.Int
+    ) -> tuple[PropagateKinematicsCarry, None]:
+
+        (i_X_0,) = carry
+
+        # Compute the base (0) to link (i) adjoint matrix.
+        # This works fine since we traverse the kinematic tree following the link
+        # indices assigned with BFS.
+        i_X_0_i = i_X_λi[i] @ i_X_0[λ[i]]
+        i_X_0 = i_X_0.at[i].set(i_X_0_i)
+
+        return (i_X_0,), None
+
+    (i_X_0,), _ = jax.lax.scan(
+        f=propagate_kinematics,
+        init=propagate_kinematics_carry,
+        xs=np.arange(start=1, stop=model.number_of_links()),
+    )
+
+    # ============================
+    # Compute doubly-left Jacobian
+    # ============================
+
+    J = jnp.zeros(shape=(6, 6 + model.dofs()))
+
+    Jb = i_X_0[link_index]
+    J = J.at[0:6, 0:6].set(Jb)
+
+    # To make JIT happy, we operate on a boolean version of κ(i).
+    # Checking if j ∈ κ(i) is equivalent to: κ_bool(j) is True.
+    κ_bool = model.kin_dyn_parameters.support_body_array_bool[link_index]
+
+    def compute_jacobian(J: jtp.MatrixJax, i: jtp.Int) -> tuple[jtp.MatrixJax, None]:
+
+        def update_jacobian(J: jtp.MatrixJax, i: jtp.Int) -> jtp.MatrixJax:
+
+            ii = i - 1
+
+            Js_i = i_X_0[link_index] @ Adjoint.inverse(i_X_0[i]) @ S[i]
+            J = J.at[0:6, 6 + ii].set(Js_i.squeeze())
+
+            return J
+
+        J = jax.lax.select(
+            pred=κ_bool[i],
+            on_true=update_jacobian(J, i),
+            on_false=J,
+        )
+
+        return J, None
+
+    L_J_WL_B, _ = jax.lax.scan(
+        f=compute_jacobian,
+        init=J,
+        xs=np.arange(start=1, stop=model.number_of_links()),
+    )
+
+    return L_J_WL_B
+
+
+@jax.jit
+def jacobian_full_doubly_left(
+    model: js.model.JaxSimModel,
+    *,
+    joint_positions: jtp.VectorLike,
+) -> tuple[jtp.Matrix, jtp.Array]:
+    r"""
+    Compute the doubly-left full free-floating Jacobian of a model.
+
+    The full Jacobian is a 6x(6+n) matrix with all the columns filled.
+    It is useful to run the algorithm once, and then extract the link Jacobian by
+    filtering the columns of the full Jacobian using the support parent array
+    :math:`\kappa(i)` of the link.
+
+    Args:
+        model: The model to consider.
+        joint_positions: The positions of the joints.
+
+    Returns:
+        The doubly-left full free-floating Jacobian of a model.
+    """
+
+    _, _, s, _, _, _, _, _, _, _ = utils.process_inputs(
+        model=model, joint_positions=joint_positions
+    )
+
+    # Get the parent array λ(i).
+    # Note: λ(0) must not be used, it's initialized to -1.
+    λ = model.kin_dyn_parameters.parent_array
+
+    # Compute the parent-to-child adjoints and the motion subspaces of the joints.
+    # These transforms define the relative kinematics of the entire model, including
+    # the base transform for both floating-base and fixed-base models.
+    i_X_λi, S = model.kin_dyn_parameters.joint_transforms_and_motion_subspaces(
+        joint_positions=s, base_transform=jnp.eye(4)
+    )
+
+    # Allocate the buffer of transforms base -> link.
+    B_X_i = jnp.zeros(shape=(model.number_of_links(), 6, 6))
+    B_X_i = B_X_i.at[0].set(jnp.eye(6))
+
+    # =============================
+    # Compute doubly-left Jacobian
+    # =============================
+
+    # Allocate the Jacobian matrix.
+    # The Jbb section of the doubly-left Jacobian is an identity matrix.
+    J = jnp.zeros(shape=(6, 6 + model.dofs()))
+    J = J.at[0:6, 0:6].set(jnp.eye(6))
+
+    ComputeFullJacobianCarry = tuple[jtp.MatrixJax, jtp.MatrixJax]
+    compute_full_jacobian_carry: ComputeFullJacobianCarry = (B_X_i, J)
+
+    def compute_full_jacobian(
+        carry: ComputeFullJacobianCarry, i: jtp.Int
+    ) -> tuple[ComputeFullJacobianCarry, None]:
+
+        ii = i - 1
+        B_X_i, J = carry
+
+        # Compute the base (0) to link (i) adjoint matrix.
+        B_Xi_i = B_X_i[λ[i]] @ Adjoint.inverse(i_X_λi[i])
+        B_X_i = B_X_i.at[i].set(B_Xi_i)
+
+        # Compute the ii-th column of the B_S_BL(s) matrix.
+        B_Sii_BL = B_Xi_i @ S[i]
+        J = J.at[0:6, 6 + ii].set(B_Sii_BL.squeeze())
+
+        return (B_X_i, J), None
+
+    (B_X_i, J), _ = jax.lax.scan(
+        f=compute_full_jacobian,
+        init=compute_full_jacobian_carry,
+        xs=np.arange(start=1, stop=model.number_of_links()),
+    )
+
+    # Convert adjoints to SE(3) transforms.
+    # Returning them here prevents calling FK in case the output representation
+    # of the Jacobian needs to be changed.
+    B_H_L = jax.vmap(lambda B_X_L: Adjoint.to_transform(B_X_L))(B_X_i)
+
+    # Adjust shape of doubly-left free-floating full Jacobian.
+    B_J_full_WL_B = J.squeeze().astype(float)
+
+    return B_J_full_WL_B, B_H_L

jaxsim/rbda/rnea.py

@@ -0,0 +1,237 @@
+from typing import Tuple
+
+import jax
+import jax.numpy as jnp
+import jaxlie
+
+import jaxsim.api as js
+import jaxsim.typing as jtp
+from jaxsim.math import Adjoint, Cross, Quaternion, StandardGravity
+
+from . import utils
+
+
+def rnea(
+    model: js.model.JaxSimModel,
+    *,
+    base_position: jtp.Vector,
+    base_quaternion: jtp.Vector,
+    joint_positions: jtp.Vector,
+    base_linear_velocity: jtp.Vector,
+    base_angular_velocity: jtp.Vector,
+    joint_velocities: jtp.Vector,
+    base_linear_acceleration: jtp.Vector | None = None,
+    base_angular_acceleration: jtp.Vector | None = None,
+    joint_accelerations: jtp.Vector | None = None,
+    link_forces: jtp.Matrix | None = None,
+    standard_gravity: jtp.FloatLike = StandardGravity,
+) -> Tuple[jtp.Vector, jtp.Vector]:
+    """
+    Compute inverse dynamics using the Recursive Newton-Euler Algorithm (RNEA).
+
+    Args:
+        model: The model to consider.
+        base_position: The position of the base link.
+        base_quaternion: The quaternion of the base link.
+        joint_positions: The positions of the joints.
+        base_linear_velocity:
+            The linear velocity of the base link in inertial-fixed representation.
+        base_angular_velocity:
+            The angular velocity of the base link in inertial-fixed representation.
+        joint_velocities: The velocities of the joints.
+        base_linear_acceleration:
+            The linear acceleration of the base link in inertial-fixed representation.
+        base_angular_acceleration:
+            The angular acceleration of the base link in inertial-fixed representation.
+        joint_accelerations: The accelerations of the joints.
+        link_forces:
+            The forces applied to the links expressed in the world frame.
+        standard_gravity: The standard gravity constant.
+
+    Returns:
+        A tuple containing the 6D force applied to the base link expressed in the
+        world frame and the joint forces that, when applied respectively to the base
+        link and joints, produce the given base and joint accelerations.
+    """
+
+    W_p_B, W_Q_B, s, W_v_WB, ṡ, W_v̇_WB, s̈, _, W_f, W_g = utils.process_inputs(
+        model=model,
+        base_position=base_position,
+        base_quaternion=base_quaternion,
+        joint_positions=joint_positions,
+        base_linear_velocity=base_linear_velocity,
+        base_angular_velocity=base_angular_velocity,
+        joint_velocities=joint_velocities,
+        base_linear_acceleration=base_linear_acceleration,
+        base_angular_acceleration=base_angular_acceleration,
+        joint_accelerations=joint_accelerations,
+        link_forces=link_forces,
+        standard_gravity=standard_gravity,
+    )
+
+    W_g = jnp.atleast_2d(W_g).T
+    W_v_WB = jnp.atleast_2d(W_v_WB).T
+    W_v̇_WB = jnp.atleast_2d(W_v̇_WB).T
+
+    # Get the 6D spatial inertia matrices of all links.
+    M = js.model.link_spatial_inertia_matrices(model=model)
+
+    # Get the parent array λ(i).
+    # Note: λ(0) must not be used, it's initialized to -1.
+    λ = model.kin_dyn_parameters.parent_array
+
+    # Compute the base transform.
+    W_H_B = jaxlie.SE3.from_rotation_and_translation(
+        rotation=jaxlie.SO3.from_quaternion_xyzw(xyzw=Quaternion.to_xyzw(wxyz=W_Q_B)),
+        translation=W_p_B,
+    )
+
+    # Compute 6D transforms of the base velocity.
+    W_X_B = W_H_B.adjoint()
+    B_X_W = W_H_B.inverse().adjoint()
+
+    # Compute the parent-to-child adjoints and the motion subspaces of the joints.
+    # These transforms define the relative kinematics of the entire model, including
+    # the base transform for both floating-base and fixed-base models.
+    i_X_λi, S = model.kin_dyn_parameters.joint_transforms_and_motion_subspaces(
+        joint_positions=s, base_transform=W_H_B.as_matrix()
+    )
+
+    # Allocate buffers.
+    v = jnp.zeros(shape=(model.number_of_links(), 6, 1))
+    a = jnp.zeros(shape=(model.number_of_links(), 6, 1))
+    f = jnp.zeros(shape=(model.number_of_links(), 6, 1))
+
+    # Allocate the buffer of transforms link -> base.
+    i_X_0 = jnp.zeros(shape=(model.number_of_links(), 6, 6))
+    i_X_0 = i_X_0.at[0].set(jnp.eye(6))
+
+    # Initialize the acceleration of the base link.
+    a_0 = -B_X_W @ W_g
+    a = a.at[0].set(a_0)
+
+    if model.floating_base():
+
+        # Base velocity v₀ in body-fixed representation.
+        v_0 = B_X_W @ W_v_WB
+        v = v.at[0].set(v_0)
+
+        # Base acceleration a₀ in body-fixed representation w/o gravity.
+        a_0 = B_X_W @ (W_v̇_WB - W_g)
+        a = a.at[0].set(a_0)
+
+        # Force applied to the base link that produce the base acceleration w/o gravity.
+        f_0 = (
+            M[0] @ a[0]
+            + Cross.vx_star(v[0]) @ M[0] @ v[0]
+            - W_X_B.T @ jnp.vstack(W_f[0])
+        )
+        f = f.at[0].set(f_0)
+
+    # ======
+    # Pass 1
+    # ======
+
+    ForwardPassCarry = Tuple[jtp.MatrixJax, jtp.MatrixJax, jtp.MatrixJax, jtp.MatrixJax]
+    forward_pass_carry: ForwardPassCarry = (v, a, i_X_0, f)
+
+    def forward_pass(
+        carry: ForwardPassCarry, i: jtp.Int
+    ) -> Tuple[ForwardPassCarry, None]:
+
+        ii = i - 1
+        v, a, i_X_0, f = carry
+
+        # Project the joint velocity into its motion subspace.
+        vJ = S[i] * ṡ[ii]
+
+        # Propagate the link velocity.
+        v_i = i_X_λi[i] @ v[λ[i]] + vJ
+        v = v.at[i].set(v_i)
+
+        # Propagate the link acceleration.
+        a_i = i_X_λi[i] @ a[λ[i]] + S[i] * s̈[ii] + Cross.vx(v[i]) @ vJ
+        a = a.at[i].set(a_i)
+
+        # Compute the link-to-base transform.
+        i_X_0_i = i_X_λi[i] @ i_X_0[λ[i]]
+        i_X_0 = i_X_0.at[i].set(i_X_0_i)
+
+        # Compute link-to-world transform for the 6D force.
+        i_Xf_W = Adjoint.inverse(i_X_0[i] @ B_X_W).T
+
+        # Compute the force acting on the link.
+        f_i = (
+            M[i] @ a[i]
+            + Cross.vx_star(v[i]) @ M[i] @ v[i]
+            - i_Xf_W @ jnp.vstack(W_f[i])
+        )
+        f = f.at[i].set(f_i)
+
+        return (v, a, i_X_0, f), None
+
+    (v, a, i_X_0, f), _ = (
+        jax.lax.scan(
+            f=forward_pass,
+            init=forward_pass_carry,
+            xs=jnp.arange(start=1, stop=model.number_of_links()),
+        )
+        if model.number_of_links() > 1
+        else [(v, a, i_X_0, f), None]
+    )
+
+    # ======
+    # Pass 2
+    # ======
+
+    τ = jnp.zeros_like(s)
+
+    BackwardPassCarry = Tuple[jtp.MatrixJax, jtp.MatrixJax]
+    backward_pass_carry: BackwardPassCarry = (τ, f)
+
+    def backward_pass(
+        carry: BackwardPassCarry, i: jtp.Int
+    ) -> Tuple[BackwardPassCarry, None]:
+
+        ii = i - 1
+        τ, f = carry
+
+        # Project the 6D force to the DoF of the joint.
+        τ_i = S[i].T @ f[i]
+        τ = τ.at[ii].set(τ_i.squeeze())
+
+        # Propagate the force to the parent link.
+        def update_f(f: jtp.MatrixJax) -> jtp.MatrixJax:
+
+            f_λi = f[λ[i]] + i_X_λi[i].T @ f[i]
+            f = f.at[λ[i]].set(f_λi)
+
+            return f
+
+        f = jax.lax.cond(
+            pred=jnp.logical_or(λ[i] != 0, model.floating_base()),
+            true_fun=update_f,
+            false_fun=lambda f: f,
+            operand=f,
+        )
+
+        return (τ, f), None
+
+    (τ, f), _ = (
+        jax.lax.scan(
+            f=backward_pass,
+            init=backward_pass_carry,
+            xs=jnp.flip(jnp.arange(start=1, stop=model.number_of_links())),
+        )
+        if model.number_of_links() > 1
+        else [(τ, f), None]
+    )
+
+    # ==============
+    # Adjust outputs
+    # ==============
+
+    # Express the base 6D force in the world frame.
+    W_f0 = B_X_W.T @ f[0]
+
+    return W_f0.squeeze(), jnp.atleast_1d(τ.squeeze())