Self-driving laboratories are increasingly presented as a route to faster, more systematic discovery by coupling algorithmic experiment selection with automated synthesis, testing, and analysis. Yet the same closed-loop structure that can improve experimental efficiency can also repeat hidden biases, amplify measurement error, misinterpret model confidence, and convert poorly specified objectives into reproducible but scientifically weak decisions. This realist review asks not whether laboratory autonomy works in the abstract, but what works, for whom, under which technical and organizational conditions, through which mechanisms, and with what outcomes. The review develops an initial programme theory, searches iteratively for explanatory evidence, and organizes findings as context–mechanism–outcome configurations across design, synthesis, testing, analysis, uncertainty management, and pharmaceutical translation. The synthesis identifies four conditional acceleration mechanisms: faster feedback between prediction and measurement, information-efficient experiment selection, consistent execution of supported procedures, and cumulative learning from structured experimental records. It also identifies four counter-mechanisms: reinforcement of biased or narrow data, exploitation of measurement noise and proxy objectives, propagation of miscalibrated uncertainty, and loss of reproducibility through brittle interfaces or underspecified protocols. The central contribution is a proposed staged-autonomy interpretation in which decision authority expands only when the relevant objective, assay, instrumentation, data lineage, uncertainty estimates, failure recovery, and human escalation pathways have demonstrated adequate reliability. The available evidence remains strongest for bounded chemical and materials tasks and substantially weaker for routine pharmaceutical decision-making, biological complexity, independent replication, and long-duration operation. Self-driving laboratories may accelerate discovery, but acceleration is not an intrinsic property of autonomy; it is an outcome produced only when enabling contexts activate reliable scientific mechanisms while safeguards interrupt error-amplifying feedback.