POMDP Coder

Induces a reusable symbolic task interface before any model program is proposed.

Converts one annotated episode into symbolic transition records aligned with the induced task interface.

Learns the initial-state sampler.

Learns the generative transition model.

Learns the symbolic observation model.

Learns the reward model.

Repairs the initial-state sampler.

Repairs the generative transition model.

Repairs the symbolic observation model.

Repairs the reward model.

Translates a natural-language instruction into a machine-checkable goal DSL used by the runtime planner.

Asks a multimodal LLM to judge which candidate observable atoms are visually supported at the current step.