EchoGen

Safe image-to-physics prototyping with a validated simulation DSL.

Purpose

EchoGen explores how multiple vision-language-model samples can be aggregated into a simplified 2D physics scene without executing model-generated code.

Safety and validation

  • AST parsing with a small allowlist of simulation operations.
  • Finite numeric and geometry checks.
  • Deterministic one-to-one matching for repeated objects across model samples.
  • Fail-closed handling for ambiguous or distant matches.
  • Offline fixtures, packaged sample assets, and Python 3.9/3.11 CI.

This is a research prototype, not a physically accurate predictor.

View the repository