SBL/Self-Building Labs

Self-Building Labs

From scientific goals to new laboratory capabilities.

Jiaru Bai†Yuchi Zhao†Han HaoMiroslav BogdanovicYang Cao*Varinia Bernales*Alán Aspuru-Guzik*

† Equal contribution* Corresponding authors

The Matter Lab @ University of Toronto · Acceleration Consortium · Vector Institute · NVIDIA

Paper PDF arXiv

A lab that builds
the capabilities science needs.

Self-Building Labs build and retain the controls, skills, workflows, and layouts that new experiments need.

  1. Identify what is missing

    Find the controls, operations, and procedures the lab needs.

  2. Build the capability

    Build digital twins, interfaces, and robot skills. Simulate operations before approved physical trials.

  3. Validate and reuse

    Refine operations with physical feedback. Save successful skills with clear reuse conditions.

How SBL changes laboratory automation.

Self-driving labs run and optimize experiments using existing capabilities. SBL develops new ones, starting from a manual lab or SDL.

CapabilityTraditional labSelf-driving labSelf-building lab
Running experimentsResearcher-led proceduresAutomation with existing capabilitiesWorkflows with capabilities built as needed
Adding capabilitiesResearcher developmentUsually specialist engineeringAgent development and validation; human oversight
Retaining knowledgeProcedures and experienceSetup-specific protocols and scriptsReusable skills with validation and use conditions

Researchers set goals and resolve uncertainties. SBL develops the capabilities to act.

Start with what you have.

Reuse what the lab has. Build what the next experiment needs.

An empty laboratory

Plan the layout, twin, skills, and workflows around the scientific goal.

A manual laboratory

Add a twin, controls, skills, and workflows to existing instruments.

A lab with a new layout

Rearrange existing equipment, update the twin, and adapt the skills.

An SDL with a new protocol

Reuse equipment and skills. Build the new workflow and missing capabilities.

From a real lab
to reusable capabilities.

Model equipment from images and documents. Develop skills through simulation and physical feedback.

Device interfaces

SBL built a joint-servoing interface for the robot arm.

Feedback-guided skills

Force and torque feedback guided fine insertion beyond camera and calibration accuracy.

Executable workflows

Combine controls, manipulation, and procedures into reusable workflows. Return control to the agent when operating conditions fail.

Explore the xArm and six-vial preparation workflows in 3D.

Drag to orbit · Scroll to zoom

From a scientific goal
to a physical experiment.

Quantum-dot synthesis

For CdS synthesis, SBL developed liquid handling and coordinated transfer, sealing, regrasping, and optical measurement.

Researchers set goals and approved plans. Execution used runtime guardrails and sensory feedback.

96
reaction wells
19
robotic stages
8 h
at 60 °C
One complete experiment · 4× speed

See the experimental results.

CdS–thiourea quantum dots
Emission 560 nm · Excitation 360 nm

Explore the 96-well pH and reagent screen: select a well and timepoint.

Well map

560 nm

0 RFU

Select a well to inspect its fluorescence.

Well H8

Fluorescence at 560 nm

Incubator temperature

60 °C setpoint · Gaps mark unlogged periods

What self-building labs will unlock in the future.

The goal is lasting capabilities that other labs can reuse.

Less rebuilding

Test whether reusing validated skills improves performance and reduces engineering effort and human intervention.

Skills shared across labs

Share twins, controls, skills, records, and validation conditions. A capability map could match goals to equipment and gaps. Check transfers locally.

Science with limited support

Build on available equipment in field stations, hazardous sites, or space, despite limited resources, communication, and support.

Acknowledgements

J.B. acknowledges funding from the Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship Program, a program by Schmidt Futures. A.A.-G. thanks Anders G. Frøseth for his generous support and acknowledges the generous support of Natural Resources Canada and the Canada 150 Research Chairs program.

This work was supported by the AI2050 program of Schmidt Sciences. This research is part of the University of Toronto’s Acceleration Consortium, which receives funding from the CFREF-2022-00042 Canada First Research Excellence Fund. This work was supported by the Defense Advanced Research Projects Agency (DARPA) under Agreement No. HR0011262E022.

Citation

@misc{bai2026selfbuildinglabs,
      title={Self-Building Labs},
      author={Jiaru Bai and Yuchi Zhao and Han Hao and Miroslav Bogdanovic and Yang Cao and Varinia Bernales and Alán Aspuru-Guzik},
      year={2026},
      eprint={2610.23826},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.23826},
}