Advanced optimization

CMAESOptimizer keeps the upstream normalized [-1, 1] search space and parameter layout. The objective expects raw simulated joint positions and encoder-frame measured positions; it applies the candidate encoder bias exactly once internally.

The historical call remains valid after explicitly binding a manually-created environment (the packaged scripts already do this):

from pace_sim2real.utils import bind_environment

bind_environment(env)
optimizer.update_simulator(robot, joint_ids, initial_encoder_position)

For new code, the explicit form is clearer and needs no binding registry:

optimizer.update_simulator(env, robot, joint_ids, initial_encoder_position)

Subclass CMAESOptimizer.tell() for a weighted or velocity-aware loss, and add a known-parameter recovery test before trusting a new objective.

tell() requires one finite floating-point [population_size, joint_count] sample for both simulation and measurement. Invalid samples are rejected before they alter CMA-ES state. Bounds must be finite and ordered, and the final delay bound must be non-negative.