<p>Self-driving laboratories accelerate materials discovery by autonomously designing and executing experiments through closed-loop integration of robotics and artificial intelligence. Active learning with Gaussian processes has enabled efficient phase diagram mapping, reducing required measurements by approximately 80% compared to conventional grid sampling. However, current approaches treat each system independently, discarding accumulated knowledge despite systematic similarities across related materials families. Here we introduce PhaseTransfer, a transfer learning framework that leverages previously characterized phase diagrams to accelerate mapping of new systems. PhaseTransfer combines a target model trained on current data with source models from related diagrams using spatially varying reliability assessments and focusing sampling on regions where transferred knowledge proves unreliable. Validation across synthetic benchmarks and experimental implementation on an autonomous microfluidic platform for biological condensate screening demonstrates an order of 50% reduction in sampling requirements compared to conventional active learning. By enabling knowledge reuse across investigations, transfer learning substantially enhances both the efficiency and generality of autonomous materials discovery.</p>