Source-linked AI summary
Self-Building Labs
Jiaru Bai, Yuchi Zhao, Han Hao, Miroslav Bogdanovic, Yang Cao, Varinia Bernales, Alán Aspuru-Guzik
TL;DR
Self-driving laboratories are limited to the procedures and capabilities established when they are configured, and adapting them to new experiments requires substantial engineering. SBL identifies and develops missing laboratory capabilities from a scientific goal, then demonstrates physical workflow execution and simulated coordination across robots and workstations.
Problem
Self-driving laboratories are bounded by their established equipment and operational capabilities, while adapting them to new research questions typically requires time-consuming specialist engineering.
Source: §1 · p. 1Method
SBL takes scientific objectives and a description of the current laboratory, then identifies gaps and develops missing layouts, digital twins, robot skills, and experimental workflows.
Source: §1 · p. 2, §2.1 · p. 3Results
Takeaways & Limitations
The quantum-dot workflow demonstrates capability development in a physical laboratory, while coordination across further robots and workstations was examined in simulation.
Source: §3 · p. 9Interpreting and weighing sensor sources still required human expertise, and coordination across further robots and workstations was examined in simulation.
Source: §3 · p. 9
Abstract
from arXiv · showhide
Self-driving laboratories (SDLs) accelerate scientific discovery by automating experimental execution and optimization. AI agents can increasingly propose new experiments, but existing SDLs can only run the procedures they were built for. Adapting them to a new experiment still takes extensive manual engineering, which makes laboratory flexibility the bottleneck for scaling discovery. We introduce the self-building laboratory (SBL), which designs, configures, and operates its own laboratory setup given a high-level scientific goal. SBL starts from whatever already exists, from an empty room to a working SDL, and builds the missing capabilities: the lab layout, its digital twin, robot skills, and experimental workflows. We demonstrate the system working across different robots and setups, scaling from a single workstation to laboratories with several robots working together. We validate SBL's end-to-end capabilities through a real-world quantum-dot synthesis experiment. By shifting the role of human researchers from manually configuring and programming device automation to specifying scientific objectives, SBL makes scientific automation more accessible, flexible, and scalable. Project website: https://self-building-labs.github.io/
1 Introduction
Existing self-driving laboratories are bounded by the equipment and operational capabilities set during development, while adapting them to new questions often requires time-consuming specialist engineering. SBL brings the identification, development, and validation of missing capabilities into the scientific process, from digital-twin construction to robot skills and workflows.
1 Introduction
- SDL autonomy is bounded by the equipment and operational capabilities established during development, limiting what procedures a platform can run.New research questions may require unanticipated procedures or instruments, whose integration typically requires specialists and time-consuming hardware and software work.Source: §1 · pp. 1–2
- Recent advances in agents, simulation, and mobile manipulation support flexible experimental execution, but adapting laboratories to new scientific requirements still takes substantial time and expertise.These advances include agents that plan experiments and generate instrument-control code, physics-based digital twins, and robots that transfer samples between instruments.Source: §1 · p. 2
- SBL incorporates the identification, development, and validation of task-required operational capabilities into the autonomous scientific process.Researchers specify scientific objectives; the system identifies capability gaps, develops and checks candidate operations in a digital twin, then refines and validates them through physical interaction.Source: §1 · p. 2
- SBL develops task-relevant digital twins, execution interfaces, and reusable sensory-feedback skills, then combines them into workflows for physical execution with existing instruments.Its capability development spans digital-twin construction, operational skills, experimental workflow building, and laboratory expansion.Source: §1 · p. 3
- SBL incorporates capability development into the scientific process, while simulation also demonstrates coordination across different robotic configurations and interconnected workstations.The authors suggest this could reduce engineering effort and make laboratory automation more accessible to researchers without specialized automation expertise.Source: §1 · p. 3
2 Results
SBL builds task-specific laboratory capabilities from the user’s goal and the laboratory’s current state, linking digital-twin construction, skill development, safety, and workflow execution. Demonstrations include CdS quantum-dot synthesis and coordinated operation across different robots and workstations.
2.1 Overview
- SBL starts from a user’s scientific goal and laboratory description, then builds the missing layout, digital twin, robot skills, and experimental workflows, with user approval for real-world trial and error.The starting point can range from an empty room to a working SDL.Source: §2.1 · pp. 3–4, Table 2 · p. 4
2.2 Building digital twin
- The twin agent reconstructs equipment and calibrates its spatial arrangement using robot-mounted camera observations, then supports simulation and physical motion planning through corresponding device interfaces.Observed and rendered views are compared iteratively to update equipment poses.Source: §2.2 · p. 4, Figure 3 · p. 5
2.3 Building laboratory skills
- Scientific requirements drive development of device interfaces and embodied skills; simulation explores action sequences, while sensory feedback refines precision-sensitive interactions on the real robot.Successful, repeatable procedures are encoded as executable skills with sensor conditions and recovery branches, allowing nominal operation without agent involvement.Source: §2.3 · pp. 4, 6
- SBL explores hazardous scenarios in simulation to identify risks and generate candidate safety guidelines, while physical execution uses operator stops, speed limits, and force-sensor monitoring.Unexpected contact can trigger protective stops.Source: §2.3 · pp. 6–7
2.4 Building experimental workflow
- For the quantum-dot workflow, SBL executed a 236-command OT-2 protocol and robotic plate transfers involving sealing, re-grasping, and reader loading.The protocol was simulator-checked and hash-frozen before execution; the arm used an intermediate holder to re-grasp the plate for reader insertion.Source: §2.4 · p. 7, Figure 5 · p. 8
- At high pH, CdS quantum-dot formation was preferential, with optimal ligand loading of 0.75–1.5 eq; excess ligand inhibits growth, while ligands stabilize soluble QDs after nucleation.The screen varied Cd:MPA ratio, Cd:thiourea ratio, and initial buffer pH at fixed Cd concentration.Source: §2.4 · p. 7
2.5 Building scalable labs
- SBL demonstrated scalability from single-arm and bimanual systems to multi-station laboratories, coordinating mobile robots to transfer materials between workstations.Demonstrations included a UR3–UR5 handoff around a shelf and a buffer-preparation workflow linking UR-based solid handling with xArm-based liquid handling.Source: §2.5 · p. 9
3 Discussion
SBL incorporates development and validation of experimental capabilities into scientific work, but what it can build depends on what laboratory infrastructure makes observable and verifiable. Its safety and long-term reuse also depend on physical checks, human expertise, and validation under changed conditions.
3 Discussion
- SBL develops missing control interfaces and success conditions, demonstrating both in the physical quantum-dot workflow while examining coordination across further robots and workstations in simulation.Examples include joint-servoing control and feedback-guided insertion when scene calibration was insufficiently precise.Source: §3 · p. 9
- The agent’s capabilities are bounded by what laboratory infrastructure lets it observe and verify.Source: §3 · p. 9
- Twin-based safety checks cover represented geometry, while chemical hazards and uncertain contact conditions require additional checks during physical execution.Safe capability development also requires an exploration space, enforced limits on physical actions, and execution observations to refine the twin and its constraints.Source: §3 · p. 9
- Human involvement remains necessary when new operating conditions emerge during long-horizon tasks, and researchers supply physical facts unavailable through interfaces.Assumptions needing clarification, overly cautious motion, and agent disagreements can require human intervention.Source: §3 · pp. 9, 11
- Figure 6 depicts a simulated layout combining solid handling, mobile manipulation, and liquid handling, with a coordinated workflow and task-relevant powder and fluid models.Source: Figure 6 · p. 10
- Successful exploration must be distilled into reusable procedures, with adapted procedures checked and validated when equipment or layouts change.The agentic cycles were too slow for the chemical workflow; reusable executable policies run without the language model and return control when conditions fall outside them.Source: §3 · p. 11
- A global capability map could connect scientific objectives to available instruments, skills, and resources, while deployment in another laboratory still requires local checks.Shareable resources include equipment models, digital-twin assets, device interfaces, operational skills, execution records, and safety guidelines with applicability and validation conditions.Source: §3 · p. 11
4 Methods
The methods describe the physical platform and software used to construct laboratory representations, generate workflows, and control equipment. They also specify the quantum-dot chemistry and measurement procedures.
4 Methods
- The experimental platform combines an xArm5 on a motorized rail, an OT-2 pipetting robot, a plate sealer, a multimode reader, and RGB-D cameras.Instrument services expose authenticated HTTP endpoints, status access, OpenAPI schemas, and function documentation.Source: §4 · p. 11
- GPT-6-Astra constructs laboratory representations and workflows, while NVIDIA Isaac Sim implements the digital twin using CAD models, reconstructed geometry, photographs, and device documentation.Source: §4 · p. 11
- The twin tracks articulated joints and liquid states, while cuRobo plans motion with collision representations and the workflow records commands, device status, measurements, and run metadata.Agents access instruments through REST and approved MCP tools; REST authorization and device claims are recorded.Source: §4 · p. 12
- Quantum-dot synthesis uses prepared buffer and reagent stocks, with absorption measured at 350, 400, and 450 nm and emission recorded from 450–680 nm.Emission measurements use 360 nm excitation at 10 nm intervals; stock solutions are prepared in a fumehood with appropriate PPE.Source: §4 · p. 12