One prioritized descent of a digit's fit: a primary residual (where the
point goes) met first, and a secondary one (how it is turned, how its
joints couple) bettered only in what the primary leaves free, inside the
range the projection keeps.
Each step is the damped least-squares step of the primary task, h1 =
-(J1'J1 + mu I)^-1 J1'r1, plus the secondary task's damped step taken in
the regularized null complement, N = I - (J1'J1 + mu I)^-1 J1'J1. This
damped complement is not an exact null projector; actual primary admission
below keeps the task priority. The direction follows
the task-priority inverse kinematics of Nakamura, Hanafusa and Yoshikawa
1987 ("Task-priority based redundancy control of robot manipulators",
IJRR 6(2):3-15) in the damped form of Chiaverini 1997 ("Singularity-robust
task-priority redundancy resolution", IEEE TRA 13(3):398-410). A variable
standing on a bound its descent pushes against (the projection undoes a
resolution's move along the negative gradient) is held for the step, the
active set of a projected Newton method (Bertsekas 1982); the gradient is
the primary's while the primary is unmet, so the secondary never holds a
joint the primary needs.
A trial is accepted lexicographically: while the primary is unmet, only
when it lowers the primary; once met (within the effector's motion for one
resolution step of every joint, tolerance), when it keeps the primary met
and lowers the secondary. The damping follows Levenberg-Marquardt with
Nielsen's update at its bounds (Madsen, Nielsen and Tingleff 2004,
Algorithm 3.16, tau = 1e-3): a third on an accepted step, a doubling growth
on a rejected one, so the steps shrink until the descent stops, when the
step it would take moves no variable by more than step. Once the primary
is met, a rejected secondary step is first backtracked by halves under the
same projected acceptance, down to that original resolution. This admits a
feasible shorter step despite the damped complement's primary leakage.
Finite differences keep two task probes per variable; a separable clipped
coordinate uses its actual feasible span instead of dividing by 2*step,
which would halve a derivative at a range end. Interior and coupled
projections retain their existing central denominator.
Residual, projection and tolerance callbacks are pure functions of their
supplied values throughout this synchronous descent, as the finite-difference
and damping law require. Rejected damping trials keep the accepted point,
Jacobians and active set unchanged; only acceptance rebuilds linearization.
A satisfied primary with zero squared secondary cannot admit any strictly
improving trial and returns immediately. With no variables, the projected
point and empty-column primary Jacobian are the only possible result.
One prioritized descent of a digit's fit: a primary residual (where the point goes) met first, and a secondary one (how it is turned, how its joints couple) bettered only in what the primary leaves free, inside the range the projection keeps.
Each step is the damped least-squares step of the primary task, h1 = -(J1'J1 + mu I)^-1 J1'r1, plus the secondary task's damped step taken in the regularized null complement, N = I - (J1'J1 + mu I)^-1 J1'J1. This damped complement is not an exact null projector; actual primary admission below keeps the task priority. The direction follows the task-priority inverse kinematics of Nakamura, Hanafusa and Yoshikawa 1987 ("Task-priority based redundancy control of robot manipulators", IJRR 6(2):3-15) in the damped form of Chiaverini 1997 ("Singularity-robust task-priority redundancy resolution", IEEE TRA 13(3):398-410). A variable standing on a bound its descent pushes against (the projection undoes a resolution's move along the negative gradient) is held for the step, the active set of a projected Newton method (Bertsekas 1982); the gradient is the primary's while the primary is unmet, so the secondary never holds a joint the primary needs.
A trial is accepted lexicographically: while the primary is unmet, only when it lowers the primary; once met (within the effector's motion for one resolution step of every joint,
tolerance), when it keeps the primary met and lowers the secondary. The damping follows Levenberg-Marquardt with Nielsen's update at its bounds (Madsen, Nielsen and Tingleff 2004, Algorithm 3.16, tau = 1e-3): a third on an accepted step, a doubling growth on a rejected one, so the steps shrink until the descent stops, when the step it would take moves no variable by more thanstep. Once the primary is met, a rejected secondary step is first backtracked by halves under the same projected acceptance, down to that original resolution. This admits a feasible shorter step despite the damped complement's primary leakage. Finite differences keep two task probes per variable; a separable clipped coordinate uses its actual feasible span instead of dividing by 2*step, which would halve a derivative at a range end. Interior and coupled projections retain their existing central denominator. Residual, projection and tolerance callbacks are pure functions of their supplied values throughout this synchronous descent, as the finite-difference and damping law require. Rejected damping trials keep the accepted point, Jacobians and active set unchanged; only acceptance rebuilds linearization. A satisfied primary with zero squared secondary cannot admit any strictly improving trial and returns immediately. With no variables, the projected point and empty-column primary Jacobian are the only possible result.