Science-Gym: a simple testbed for AI-driven scientific discovery
What this work does
Provides seven Gym-compatible scientific simulations with separate simulation and equation-problem interfaces for data collection and interpretable equation discovery.
Connection to discovery
A training interface for separating experiment collection, control reward and equation recovery within one discovery task.
Evidence
Inspected experimental-action semantics, reset/step structure and the equation table, including the explicitly unknown ground truth for droplet friction.
Scope and limitations
Simulation access is not a live physics laboratory. The paper notes missing explicit support for discovering new objects/variables; not every task has a closed-form symbolic ground truth.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.