Questions about jacobian in littledog.ipynb

[matheecs]

Hi @RussTedrake, the littledog.ipynb is a great demo about trajectory optimization. But I have a question about the calculation of jacobian Jq_WF:

# Hdot = sum_i cross(p_FootiW-com, contact_force_i)
def angular_momentum_constraint(vars, context_index):
    q, com, Hdot, contact_force = np.split(vars, [nq, nq + 3, nq + 6])
    ......
    if isinstance(vars[0], AutoDiffXd):
            ......
            ad_p_WF = InitializeAutoDiff(
                p_WF, np.hstack((Jq_WF, np.zeros((3, 18))))
            )
            torque = torque + np.cross(
                ad_p_WF.reshape(3) - com, contact_force[:, i]
            )
    else:
            ......

Does it need to be change to Jq_WF @ ExtractGradient(q)?

# Hdot = sum_i cross(p_FootiW-com, contact_force_i)
def angular_momentum_constraint(vars, context_index):
    q, com, Hdot, contact_force = np.split(vars, [nq, nq + 3, nq + 6])
    ......
    if isinstance(vars[0], AutoDiffXd):
            ......
            ad_p_WF = InitializeAutoDiff(
                p_WF, np.hstack((Jq_WF @ ExtractGradient(q), np.zeros((3, 18))))
            )
            torque = torque + np.cross(
                ad_p_WF.reshape(3) - com, contact_force[:, i]
            )
    else:
            ......

Thanks

[RussTedrake]

You're absolutely right that I should be using the gradients passed in. Thanks!

It happens that our current snopt solver will only ever call this method with the gradients in vars trivially initialized (so ExtractGradient(q) returns the identity). I'll PR the fixed version based on your idea.