Position: Autonomous Scientific Discovery Needs Embodied Experimentation with Learnability

Autonomous scientific discovery, particularly through self-driving laboratories (SDLs) integrated with large language models, has achieved breakthroughs in materials science, drug design, and beyond. However, current implementations exhibit a critical limitation: cognitive parameters remain immutable post-training, motor behaviours adhere to pre-programmed sequences, and experimental optimisation proceeds without leveraging institutional knowledge. Thus, we propose a formalised trinity of learnability comprising cognitive design, experimentation optimisation, and embodied execution, where an evolving cognitive core serves as the foundation for discovery. This architecture synergises the cognitive core with adaptive embodied skills to translate lab-specific variances into robustness while refining active experimental designs to prune explosive search spaces. This synergy enables SDLs to foster the emergence of high-efficiency discovery. Taken together, this position paper argues that autonomous discovery requires learnability through embodied experimentation, not merely automation sophistication. As a proof of concept, we developed a bespoke embodied laboratory prototype for the Electro-Fenton task. By leveraging historical memory and adaptive skills, our agent navigated a 200-million search space to achieve 99.7% efficiency in just 15 rounds.