1 pose_generation.candidate_mixture

pose_generation.candidate_mixture

Mixed finite-candidate view generation with full-shell provenance.

The mixture wrapper keeps CandidateViewGeneratorConfig backward-compatible by instantiating one normal generator per component, overriding only the component fields, and concatenating the full sampled shells. Component counts therefore define the fixed candidate budget exactly, while each output row retains strategy_id, mixture_id, sampler_probability, and optional component name.

Ordering is part of the data contract: invalid candidates remain in the full-shell masks and provenance arrays; compact valid views are recoverable through CandidateSamplingResult.candidate_shell_indices(). Target-point components require CandidateGenerationRuntimeContext.target_center_world because the thesis V1 actor may condition on observed/predicted target records but not on GT target geometry.

1.1 Theory

The default target-conditioned mixture has 60 full-shell rows: forward_local 24, target_bearing_local 24, and lateral_target_bypass 12. Richer local-refinement, revisit-backtrack, and free-shell components are explicit ablations. Each row records stable position_id, strategy_id, mixture_id, component name, and sampler_probability = 1/N. These provenance arrays let rollout/Q_H stores audit which finite-action family produced each candidate without changing the compact valid-action contract.

1.2 Classes

Name Description
CandidateMixtureComponentConfig One fixed-count candidate-family component inside a mixture sampler.
CandidateMixtureViewGeneratorConfig Config-as-factory for fixed-count mixed candidate tables.
CandidateMixtureViewGenerator Generate a fixed-size candidate table from multiple sampling families.

1.3 Functions

Name Description
candidate_strategy_id Return the stable integer provenance id for a view-direction family.
candidate_position_id Return the stable integer provenance id for a position family.