Merge pull request ami-iit#54 from ami-iit/automatic_differentiation

First steps towards automatic differentiation of RBDAs

src/jaxsim/math/quaternion.py

@@ -33,7 +33,7 @@ def derivative(
         omega_in_body_fixed: bool = False,
         K: float = 0.1,
     ) -> jtp.Vector:
-        w = omega.squeeze()
+        ω = omega.squeeze()
         quaternion = quaternion.squeeze()
 
         def Q_body(q: jtp.Vector) -> jtp.Matrix:
@@ -67,10 +67,20 @@ def Q_inertial(q: jtp.Vector) -> jtp.Matrix:
             operand=quaternion,
         )
 
+        norm_ω = jax.lax.cond(
+            pred=ω.dot(ω) < (1e-6) ** 2,
+            true_fun=lambda _: 1e-6,
+            false_fun=lambda _: jnp.linalg.norm(ω),
+            operand=None,
+        )
+
         qd = 0.5 * (
             Q
             @ jnp.hstack(
-                [K * jnp.linalg.norm(w) * (1 - jnp.linalg.norm(quaternion)), w]
+                [
+                    K * norm_ω * (1 - jnp.linalg.norm(quaternion)),
+                    ω,
+                ]
             )
         )
 

src/jaxsim/physics/algos/soft_contacts.py

@@ -229,6 +229,10 @@ def contact_model(
         m = tangential_deformation.squeeze()
         ṁ = jnp.zeros_like(m)
 
+        # Note: all the small hardcoded tolerances in this method have been introduced
+        # to allow jax differentiating through this algorithm. They should not affect
+        # the accuracy of the simulation, although they might make it less readable.
+
         # ========================
         # Normal force computation
         # ========================
@@ -249,7 +253,11 @@ def contact_model(
 
         # Non-linear spring-damper model.
         # This is the force magnitude along the direction normal to the terrain.
-        force_normal_mag = jnp.sqrt(δ) * (K * δ + D * δ̇)
+        force_normal_mag = jax.lax.select(
+            pred=δ >= 1e-9,
+            on_true=jnp.sqrt(δ + 1e-12) * (K * δ + D * δ̇),
+            on_false=jnp.array(0.0),
+        )
 
         # Prevent negative normal forces that might occur when δ̇ is largely negative
         force_normal_mag = jnp.maximum(0.0, force_normal_mag)
@@ -304,7 +312,7 @@ def below_terrain():
                 v_tangential = W_ṗ_C - v_normal
 
                 # Compute the tangential force. If inside the friction cone, the contact
-                f_tangential = -jnp.sqrt(δ) * (K * m + D * v_tangential)
+                f_tangential = -jnp.sqrt(δ + 1e-12) * (K * m + D * v_tangential)
 
                 def sticking_contact():
                     # Sum the normal and tangential forces, and create the 6D force
@@ -319,9 +327,17 @@ def sticking_contact():
                     return CW_f, ṁ
 
                 def slipping_contact():
+                    # Clip the tangential force if too small, allowing jax to
+                    # differentiate through the norm computation
+                    f_tangential_no_nan = jax.lax.select(
+                        pred=f_tangential.dot(f_tangential) >= 1e-9**2,
+                        on_true=f_tangential,
+                        on_false=jnp.array([1e-12, 0, 0]),
+                    )
+
                     # Project the force to the friction cone boundary
                     f_tangential_projected = (μ * force_normal_mag) * (
-                        f_tangential / jnp.linalg.norm(f_tangential)
+                        f_tangential / jnp.linalg.norm(f_tangential_no_nan)
                     )
 
                     # Sum the normal and tangential forces, and create the 6D force
@@ -331,18 +347,18 @@ def slipping_contact():
                     # Correct the material deformation derivative for slipping contacts.
                     # Basically we compute ṁ such that we get `f_tangential` on the cone
                     # given the current (m, δ).
-                    ε = 1e-6
-                    α = -K * jnp.sqrt(δ)
+                    ε = 1e-9
                     δε = jnp.maximum(δ, ε)
-                    βε = -D * jnp.sqrt(δε)
-                    ṁ = (f_tangential_projected - α * m) / βε
+                    α = -K * jnp.sqrt(δε)
+                    β = -D * jnp.sqrt(δε)
+                    ṁ = (f_tangential_projected - α * m) / β
 
                     # Return the 6D force in the contact frame and
                     # the deformation derivative
                     return CW_f, ṁ
 
                 CW_f, ṁ = jax.lax.cond(
-                    pred=jnp.linalg.norm(f_tangential) > μ * force_normal_mag,
+                    pred=f_tangential.dot(f_tangential) > (μ * force_normal_mag) ** 2,
                     true_fun=lambda _: slipping_contact(),
                     false_fun=lambda _: sticking_contact(),
                     operand=None,

tests/test_ad_physics.py

@@ -0,0 +1,190 @@
+import jax.numpy as jnp
+import numpy as np
+import pytest
+from jax.test_util import check_grads
+from pytest import param as p
+
+from jaxsim.high_level.common import VelRepr
+from jaxsim.high_level.model import Model
+
+from . import utils_models, utils_rng
+from .utils_models import Robot
+
+
+@pytest.mark.parametrize(
+    "robot, vel_repr",
+    [
+        p(*[Robot.Ur10, VelRepr.Inertial], id="Ur10-Inertial"),
+        p(*[Robot.AnymalC, VelRepr.Inertial], id="AnymalC-Inertial"),
+        p(*[Robot.Cassie, VelRepr.Inertial], id="Cassie-Inertial"),
+    ],
+)
+def test_ad_physics(robot: utils_models.Robot, vel_repr: VelRepr) -> None:
+    """Unit test of the application of Automatic Differentiation on RBD algorithms."""
+
+    robot = Robot.Ur10
+    vel_repr = VelRepr.Inertial
+
+    # Initialize the gravity
+    gravity = np.array([0, 0, -10.0])
+
+    # Get the URDF of the robot
+    urdf_file_path = utils_models.ModelFactory.get_model_description(robot=robot)
+
+    # Build the high-level model
+    model = Model.build_from_model_description(
+        model_description=urdf_file_path,
+        vel_repr=vel_repr,
+        gravity=gravity,
+        is_urdf=True,
+    ).mutable(mutable=True, validate=True)
+
+    # Initialize the model with a random state
+    model.data.model_state = utils_rng.random_physics_model_state(
+        physics_model=model.physics_model
+    )
+
+    # Initialize the model with a random input
+    model.data.model_input = utils_rng.random_physics_model_input(
+        physics_model=model.physics_model
+    )
+
+    # ========================
+    # Extract state and inputs
+    # ========================
+
+    # Extract the physics model used in the low-level physics algorithms
+    physics_model = model.physics_model
+
+    # State
+    s = model.joint_positions()
+    ṡ = model.joint_velocities()
+    xfb = model.data.model_state.xfb()
+
+    # Inputs
+    f_ext = model.external_forces()
+    tau = model.joint_generalized_forces_targets()
+
+    # Perturbation used for computing finite differences
+    ε = jnp.finfo(jnp.array(0.0)).resolution ** (1 / 3)
+
+    # =====================================================
+    # Check first-order and second-order derivatives of ABA
+    # =====================================================
+
+    import jaxsim.physics.algos.aba
+
+    aba = lambda xfb, s, ṡ, tau, f_ext: jaxsim.physics.algos.aba.aba(
+        model=physics_model, xfb=xfb, q=s, qd=ṡ, tau=tau, f_ext=f_ext
+    )
+
+    check_grads(
+        f=aba,
+        args=(xfb, s, ṡ, tau, f_ext),
+        order=2,
+        modes=["rev", "fwd"],
+        eps=ε,
+    )
+
+    # ======================================================
+    # Check first-order and second-order derivatives of RNEA
+    # ======================================================
+
+    import jaxsim.physics.algos.rnea
+
+    W_v̇_WB = utils_rng.get_rng().uniform(size=6, low=-1)
+    s̈ = utils_rng.get_rng().uniform(size=physics_model.dofs(), low=-1)
+
+    rnea = lambda xfb, s, ṡ, s̈, W_v̇_WB, f_ext: jaxsim.physics.algos.rnea.rnea(
+        model=physics_model, xfb=xfb, q=s, qd=ṡ, qdd=s̈, a0fb=W_v̇_WB, f_ext=f_ext
+    )
+
+    check_grads(
+        f=rnea,
+        args=(xfb, s, ṡ, s̈, W_v̇_WB, f_ext),
+        order=2,
+        modes=["rev", "fwd"],
+        eps=ε,
+    )
+
+    # ======================================================
+    # Check first-order and second-order derivatives of CRBA
+    # ======================================================
+
+    import jaxsim.physics.algos.crba
+
+    crba = lambda s: jaxsim.physics.algos.crba.crba(model=physics_model, q=s)
+
+    check_grads(
+        f=crba,
+        args=(s,),
+        order=2,
+        modes=["rev", "fwd"],
+        eps=ε,
+    )
+
+    # ====================================================
+    # Check first-order and second-order derivatives of FK
+    # ====================================================
+
+    import jaxsim.physics.algos.forward_kinematics
+
+    fk = (
+        lambda xfb, s: jaxsim.physics.algos.forward_kinematics.forward_kinematics_model(
+            model=physics_model, xfb=xfb, q=s
+        )
+    )
+
+    check_grads(
+        f=fk,
+        args=(xfb, s),
+        order=2,
+        modes=["rev", "fwd"],
+        eps=ε,
+    )
+
+    # ==========================================================
+    # Check first-order and second-order derivatives of Jacobian
+    # ==========================================================
+
+    import jaxsim.physics.algos.jacobian
+
+    link_indices = [l.index() for l in model.links()]
+
+    jacobian = lambda s: jaxsim.physics.algos.jacobian.jacobian(
+        model=physics_model, q=s, body_index=link_indices[-1]
+    )
+
+    check_grads(
+        f=jacobian,
+        args=(s,),
+        order=2,
+        modes=["rev", "fwd"],
+        eps=ε,
+    )
+
+    # =====================================================================
+    # Check first-order and second-order derivatives of soft contacts model
+    # =====================================================================
+
+    import jaxsim.physics.algos.soft_contacts
+
+    p = utils_rng.get_rng().uniform(size=3, low=-1)
+    v = utils_rng.get_rng().uniform(size=3, low=-1)
+    m = utils_rng.get_rng().uniform(size=3, low=-1)
+
+    parameters = jaxsim.physics.algos.soft_contacts.SoftContactsParams.build(
+        K=10_000, D=20.0, mu=0.5
+    )
+
+    soft_contacts = lambda p, v, m: jaxsim.physics.algos.soft_contacts.SoftContacts(
+        parameters=parameters
+    ).contact_model(position=p, velocity=v, tangential_deformation=m)
+
+    check_grads(
+        f=soft_contacts,
+        args=(p, v, m),
+        order=2,
+        modes=["rev", "fwd"],
+        eps=ε,
+    )