The bottleneck for scaling discovery in self-driving laboratories has long been the human effort required to reconfigure hardware and software for each new experiment. Existing self-driving laboratories can only run the procedures they were originally built for; adapting them to a new scientific goal demands extensive manual engineering. This work identifies that rigidity as the primary constraint and introduces the self-building laboratory as an alternative framework. The SBL system designs, configures, and operates its own laboratory setup from a high-level scientific goal, starting from whatever already exists—an empty room, a working SDL, or anything in between. It constructs the missing capabilities: lab layout, digital twin, robot skills, and experimental workflows. The approach has been validated across different robots and setups, scaling from a single workstation to laboratories with several robots collaborating. End-to-end capabilities are demonstrated through a real-world quantum-dot synthesis experiment. By shifting the researcher's role from manually configuring and programming device automation to specifying scientific objectives, the framework makes scientific automation more accessible, flexible, and scalable.
Traditional self-driving laboratories rely on a fixed experimental infrastructure. An SDL built for organic synthesis cannot easily be repurposed for materials discovery without significant reengineering of robot trajectories, sensor interfaces, and workflow scripts. The labor required to adapt such systems scales with the complexity of the new experiment, creating a practical ceiling on throughput. This work argues that the adaptation burden, not the automation capacity itself, is what limits discovery rate. The self-building laboratory framework was developed to remove that burden. Its central proposition is that a laboratory system should be capable of examining its own state, identifying gaps between its current capabilities and the requirements of a new task, and then constructing the missing pieces autonomously.
SBL operates in three overlapping phases. In the perception phase, the system assesses the existing laboratory environment. This includes scanning physical layout, identifying attached devices, reading instrument metadata, and constructing a digital twin that captures the spatial and functional properties of the current setup. The digital twin is not merely a visualization tool; it is a structured model that the reasoning layer queries when planning what is missing. If the twins indicate that no robot arm has a gripper suitable for the target synthesis, or that no temperature controller covers the required range, the system flags those as gaps.
In the capability-synthesis phase, the system generates the missing components. Lab layout is reconfigured by computing feasible robot paths that avoid collision with existing equipment and by generating new station placements when the current arrangement cannot accommodate the planned experiment. Digital twins are extended or rebuilt from scratch if the existing model does not reflect the desired configuration. Robot skills—parameterized trajectories, grasp motions, and instrument-interaction sequences—are synthesized from high-level specifications. The system does not retrieve these skills from a fixed library; instead, it generates them using a combination of motion-planning algorithms and task-and-motion planning, ensuring that the resulting skills satisfy constraints such as collision avoidance, force limits, and instrument compatibility. Experimental workflows are compiled from these skills and from procedural steps such as reagent addition, heating cycles, and quenching protocols. The workflow representation is executable by the laboratory's runtime system, which handles real-time coordination of robot actions, sensor feedback, and safety monitors.
The third phase is operation. Once the lab layout, digital twin, robot skills, and workflows have been constructed, the SBL runtime executes the experiment from start to finish. The system monitors live sensor data, detects deviations from expected behavior, and makes real-time adjustments. If a reagent pump underdelivers, the system can pause, reorient, and retry. If a robot joint exceeds its torque limit, the system can replan the remaining trajectory. The runtime also records the full execution trace, enabling later analysis and transfer to future experiments.
The authors validate the framework through a real-world quantum-dot synthesis experiment. Quantum dots are semiconductor nanocrystals whose optical properties are determined by size, and their synthesis requires precise control of temperature, precursor injection timing, and solvent ratios. The SBL system was tasked with producing a specified quantum-dot size distribution from raw starting materials. Starting from a laboratory that already housed some automation but lacked a cohesive workflow for nanoparticle synthesis, the system built the required robot skills, reconfigured lab layout for safe handling of precursor chemicals, and compiled an experimental workflow that included sequential injection steps and temperature ramps. The resulting synthesized quantum dots exhibited the target optical characteristics, and the full execution was completed without human intervention in device setup. The authors report that the system scaled from a single robotic workstation to a configuration with three collaborative robots working on different subtasks—precursor mixing, temperature control, and post-synthesis purification—while maintaining coherent experimental oversight.
Beyond the quantum-dot demonstration, the paper describes additional validations across different laboratory configurations. In one setup, the SBL system was deployed in a chemistry laboratory equipped with a liquid-handling robot and an automated microscope. The system built skills for microscopic imaging, automatically identifying regions of interest and capturing images at specified intervals. In another setup, the system was used in a physics laboratory with a vacuum-compatible robot arm; it synthesized robot skills for manipulating viewport ports and controlling pressure regulation valves. Across all validations, the common pattern was that the system started from the existing hardware inventory, identified gaps, and generated the necessary capabilities without requiring a complete overhaul of the laboratory infrastructure.
The paper also discusses the representation formalisms that underpin SBL's reasoning. Lab layout is modeled as a spatial graph where nodes represent equipment stations and edges represent traversable paths between stations. The graph is annotated with constraints such as reachable workspace, cable length limits, and safety zones. When the system needs to reorganize the layout, it solves a constrained path-planning problem on this graph, seeking a configuration that satisfies the new experiment's spatial requirements while preserving feasible operation of existing equipment. Robot skills are represented as parameterized motion sequences augmented with procedural postconditions. A skill template describes the nominal trajectory, while a set of constraints defines valid parameter ranges. The runtime system evaluates these constraints in real time, braking or reparameterizing the motion if a constraint is breached. Experimental workflows are expressed in a directed-acyclic graph format, where nodes are atomic actions (e.g., "open valve," "set temperature," "acquire image") and edges encode ordering constraints. This representation allows the system to compile high-level objectives such as "synthesize quantum dots with target emission wavelength" into executable plans by connecting atomic actions through the workflow DAG.
One of the most computationally demanding aspects of SBL is the synthesis of robot skills. The system explores a space of possible motion parameters, evaluating each candidate against a suite of constraints. The evaluation is performed using a combination of analytic checks (e.g., workspace boundary tests) and simulated execution in the digital twin. Candidates that pass all checks are retained, and the final skill is selected based on a scoring function that balances success probability and execution time. The authors note that this search can be expensive for complex tasks, but they observe that the cost is amortized across multiple experiments that reuse the same or similar skills.
The digital twin plays a role throughout the SBL pipeline, not only during skill synthesis. During experiment execution, the twin is kept in sync with the physical world through a stream of sensor readings. When a discrepancy is detected—for instance, a temperature reading that deviates from the twin's prediction—the runtime can query the twin to hypothesize causes, such as a failing heater element or a miscalibrated sensor. These hypotheses are then tested through targeted actions, and the twin's model is updated accordingly. This closed-loop modeling means that the digital twin improves over the course of a single experiment, and its improvements persist across experiments.
From manual programming to objective specification
The human role in laboratory automation has traditionally been one of detailed programming. A researcher must write or generate robot trajectories, specify sensor triggers, and encode procedural branching logic for each new experiment. This work proposes a shift toward objective specification: the researcher defines what the experiment should achieve, and the SBL system determines how to achieve it using the available hardware. The authors describe this as a move from "how" to "what." The practical impact is that researchers with domain expertise in the scientific question but limited robotics or programming experience can still deploy automated experimentation. The framework abstracts away the details of motion planning, skill synthesis, and workflow compilation, presenting the researcher with a high-level interface for task description.
Scaling from single workstation to multi-robot laboratories
The validation experiments demonstrate that SBL can scale the number of participating robots without a proportional increase in human oversight. In the three-robot quantum-dot setup, each robot was assigned a distinct subtask, and the SBL runtime coordinated their actions through a shared temporal logic. The system handled inter-robot coordination such as ensuring that precursor injection by one robot did not conflict with temperature ramp-up by another. The authors report that the coordination layer successfully managed action interleaving, resource sharing, and conflict resolution in all tested scenarios. The scaling behavior was near-linear: adding a fourth robot for an additional purification step required only a modest expansion of the workflow DAG and no reengineering of existing skills.
Current limitations and boundaries
The authors acknowledge several limitations. The skill-synthesis search space can be large, and for some highly constrained robot configurations, the system may fail to find a feasible skill within a reasonable time budget. The digital twin's fidelity depends on the accuracy of the initial hardware descriptions; if the twin does not capture real-world constraints such as cable drag or tool-change limitations, the generated skills may fail at runtime. The workflow DAG compilation assumes that atomic actions are well-defined and that their ordering constraints are sufficient to guarantee a valid execution plan; in cases where actions have complex interdependencies that are not captured as simple ordering, the compiled workflow may deadlock. The paper also notes that the current framework does not handle the discovery of entirely new hardware—it works with the inventory that already exists, adding capabilities through software and motion planning, but it cannot, for example, design and build a new robot arm from scratch.
In summary, the self-building laboratory framework addresses the flexibility bottleneck in scientific automation by enabling laboratory systems to assess their own state, generate missing capabilities, and execute experiments from a high-level goal. The approach has been demonstrated across different robots and laboratory configurations, scaling from a single workstation to multi-robot setups, with a real-world quantum-dot synthesis as the flagship validation. By shifting the researcher's role from manual configuration to objective specification, the framework aims to make scientific automation more accessible and scalable. The released project website and code artifacts are intended to lower the barrier for other research groups to adopt and extend the approach.