Related software

S3E sits between symbolic-planning libraries, which assume the state is given, and tools for running vision-language models, which know nothing about planning problems. The closest tools are summarized below.

Planning libraries. Unified Planning and PDDLGym model, ground, simulate, or solve planning problems, but take the symbolic state as given. S3E builds on Unified Planning instead of reimplementing it: PDDL parsing and grounding are delegated to it, and estimated states convert back to Unified Planning state objects.

VLM evaluation harnesses. LMMs-Eval and VLMEvalKit run models on benchmarks and report aggregate metrics. LMMs-Eval can score answer log-likelihoods, but only inside benchmark task definitions.

Probability-of-“yes” scoring. The closest tool is t2v_metrics, the package behind VQAScore. It scores the probability of “Yes” for user-supplied images and question templates across local and API models, and targets the evaluation of text-to-visual generation. It has no predicate grounding, multi-option answer spaces, explicit “unknown” answer, or calibration.

Constrained generation. Libraries such as Outlines restrict what a model may generate, but do not report per-option probabilities.

Research systems. Systems that ground planners with VLM questions, such as TP-VQA, DKPROMPT, the VLM-as-grounder methods evaluated in ViPlan, and predicate-learning approaches, implement the idea inside experiment code tied to particular domains and benchmarks, rather than as an installable library.

None of these combines what S3E provides behind one interface: PDDL grounding, pluggable predicate-to-question translation, token-probability scoring over configurable answer spaces with an optional “unknown” option, interchangeable local and hosted model backends, and offline calibration.

The software paper cites each of these works.