newton.ik.IKOptimizerLM#

class newton.ik.IKOptimizerLM(model, n_batch, objectives, *a, **kw)[source]#

Bases: object

Levenberg-Marquardt optimizer for batched inverse kinematics.

The optimizer solves a batch of independent IK problems that share a single articulation model and objective list. Jacobians can be evaluated with IKJacobianType.AUTODIFF, IKJacobianType.ANALYTIC, or IKJacobianType.MIXED.

Parameters:

  • model (Model) – Shared articulation model.

  • n_batch (int) – Number of evaluation rows solved in parallel. This is typically n_problems * n_seeds after any sampling expansion.

  • objectives (Sequence[IKObjective]) – Ordered IK objectives applied to every batch row.

  • lambda_initial – Initial LM damping factor for each batch row.

  • jacobian_mode – Jacobian backend to use.

  • lambda_factor – Factor used to increase or decrease the damping term after each trial step.

  • lambda_min – Minimum allowed damping value.

  • lambda_max – Maximum allowed damping value.

  • rho_min – Minimum ratio of actual to predicted decrease required to accept a step.

  • problem_idx – Optional mapping from batch rows to base problem indices for per-problem objective data.

__init__(model, n_batch, objectives, lambda_initial=0.1, jacobian_mode=IKJacobianType.AUTODIFF, lambda_factor=2.0, lambda_min=1e-5, lambda_max=1e10, rho_min=1e-3, *, problem_idx=None)#
compute_costs(joint_q)#

Evaluate squared residual costs for a batch of joint configurations.

Parameters:

joint_q (array(ndim=2, dtype=float32)) – Joint coordinates to evaluate, shape [n_batch, joint_coord_count].

Returns:

Costs for each batch row, shape [n_batch].

Return type:

array(ndim=1, dtype=float32)

reset()#

Clear LM damping and accept/reject state before a new solve.

step(joint_q_in, joint_q_out, iterations=10, step_size=1.0)#

Run several LM iterations on a batch of joint configurations.

Parameters:

  • joint_q_in (array(ndim=2, dtype=float32)) – Input joint coordinates, shape [n_batch, joint_coord_count].

  • joint_q_out (array(ndim=2, dtype=float32)) – Output buffer for the optimized coordinates, shape [n_batch, joint_coord_count]. It may alias joint_q_in for in-place updates.

  • iterations (int) – Number of LM iterations to execute.

  • step_size (float) – Scalar applied to each computed update before integration.