Recent advances in large models have accelerated AI for Science toward agentic science, where AI agents can autonomously conduct discovery. Realising this vision requires a closed loop between computational reasoning and physical experimentation, making self‑driving laboratories (SDLs) essential. Existing SDLs show a paradox: integrating foundation models improves linguistic sophistication but weakens physical grounding. By treating models as disembodied planners that output symbolic workflows for separate robots, current designs block the development of embodied experimental intuition. Overcoming this paradox demands embedding learnability along three interdependent axes: cognitive learnability (what to investigate), experimental learnability (where to search), and embodied learnability (how to act). Guided by this view we introduce the Embodied Science Agent (ESA), which unifies digital intelligence and physical embodiment through three synergistic components: a Digital Agent that supplies cognitive and experimental learnability via memory‑based reasoning and Bayesian optimisation, an Embodied Agent that supplies embodied learnability via generalisable instrument control, and an Embodied Unified Interface that coordinates the two seamlessly. We validated ESA in autonomous wastewater‑treatment experiments, navigating a search space of over 200 million conditions. Across 15 continuous rounds ESA raised the final solution transmittance from 82% to 99.7%, a 98.5% reduction in residual dye concentration and near‑complete decolourisation, while its learned visuomotor policies generalised across varying instrument placements, illustrating a promising path toward next‑generation SDLs for autonomous scientific discovery.
Review