AI Labs Move From Automation to Self-Directed Discovery

Labs are learning to build themselves. Researchers in Alán Aspuru-Guzik’s group at the University of Toronto describe an agentic system that takes a high-level scientific goal and designs, configures, and operates its own laboratory setup. Give it an objective, and it works out the layout, simulation, robot skills, and experimental protocol.

That approach addresses the weak point in today’s self-driving laboratories. Existing systems automate experimental execution and optimization, and AI agents increasingly propose the next experiment—but each lab usually runs only the procedures it was built for. Adapting it to a new experiment still demands extensive manual engineering, making laboratory flexibility the real bottleneck for scaling discovery.

The system, called SBL, starts with whatever already exists, from an empty room to a working self-driving lab that needs rewiring for a new goal. It identifies the gaps between the objective and its current capabilities, then develops the interfaces, skills, and workflows required to close them.

A digital twin becomes the lab’s test suite

SBL rests on four capabilities: lab layout, digital twin, robot skills, and experimental workflows. A planning task that might otherwise end with a whiteboard and a mild disagreement becomes a structured reconstruction of the laboratory, built from images of the physical space and extended down to particle-based simulation of solid dispensing.

The twin sits at the center of the architecture. Robot skills act as operational primitives for each device, refined through planning in the twin and feedback from physical runs. Experimental workflows then chain those skills into complete, runnable experiments, creating the familiar agentic loop: plan, act, observe, revise.

The team validated SBL with a real-world cadmium sulfide quantum dot synthesis. It coordinated Opentrons liquid handling, plate sealing, and photophysics measurements into one working experiment, while the paper also showed how the system can scale from a single workstation to labs with several collaborating robots. The physical demonstration covered the quantum dot workflow; coordination across additional robots and workstations ran in simulation.

This matters because physical errors have physical costs. In a wet lab, a planning mistake consumes reagents and may produce the occasional shattered vial—an expensive way to discover that software needed another test.

A self-driving lab built from four AI-driven robots ran 96 miniaturized perovskite quantum dot experiments at a time in August 2025. Yet a March 2026 benchmark found that the best AI data agent answered only 38% of enterprise-style queries correctly. The message is not subtle: agents need controlled environments before anyone hands them expensive equipment and resources.

The same pattern is reaching factories

A mid-2026 paper on multi-agent AI for materials labs makes the matching case for proving agents in a digital sandbox before they touch real tools. SBL builds that sandbox as part of the job, suggesting a practical development model: treat skills as a compounding library, make the twin the test suite, and scope permits one workflow at a time.

The wider movement is called physical AI—systems connected to cameras, sensors, or robotic arms and legs that do not merely describe an action but perform it. At its CES 2026 keynote, Nvidia unveiled new open AI models designed to help machines perceive, understand, and act on the world.

AI for biology faces its own limits. At the Next Endeavor event in Mountain View, Lisa Dolan interviewed Surya Ganguli, Krithik Ramesh, and Ali Khalatpour about applying AI to biology. Ramesh noted that pre-training on proteins does not easily characterize sequences, while Ganguli said biology is harder than math because of noisy data, incomplete observations, and absent theoretical frameworks. Researchers have also used AI to model the brain by perturbing the system to simulate disorders of consciousness, and Khalatpour discussed developing space lasers with NASA funding while overcoming GPU data bottlenecks.

Industrial systems are adopting a related idea through living digital twins. These combine a dynamic, high-fidelity representation of complex operations with the ability to intervene through operations and control systems. Rob McGreevy, chief product officer at AVEVA, likens AI-infused twins to organisms that grow and evolve: adaptive, evergreen models of plant assets and processes.

They connect engineering and design information with real-time operational data through a digital thread, while sensors stream telemetry from IoT devices and industrial control equipment. Predictive and physics-based models add insight into how machines and processes might behave, allowing agentic AI to move beyond diagnostics and suggestions toward action.

The numbers show both demand and caution. A 2026 Manufacturing Leadership Council survey found that 66% of manufacturers use or plan to use agentic AI, and 40% expect some degree of factory autonomy by 2030. Deloitte’s 2025 smart manufacturing survey found that 65% of executives ranked operational risk, including business disruption, among their top concerns.

Examples already exist. At Saudi Aramco’s Yanbu refinery, optimization algorithms adjust production in a closed loop, with operators overseeing as needed; at Ras Tanura, algorithms recommend changes and operators decide whether to implement them. Nestlé used AI optimization at Nesquik plants to tune industrial dryers, reducing energy use by 16% while improving quality and yield. Thailand’s SCG Chemicals reports 99% plant reliability and a ninefold return on investment within six months from an AI-supported digital twin.

Human oversight remains the dividing line between useful autonomy and expensive chaos. The EU AI Act classifies certain AI safety components used in critical infrastructure as high-risk and requires human oversight, while the National Laboratory of the Rockies built a generative AI model to support grid operators with renewable energy forecasting and estimates of grid conditions.

McGreevy says manufacturing and critical infrastructure have operated as siloed systems for 50 years. Living digital twins aim to connect them across the value chain, helping companies respond to supply chain changes, demand, tariffs, and material availability. The machines may get the autonomy, but humans still define what healthy looks like—and decide when the system has stopped meeting the brief.

Stay connected via Google News

Follow us for the latest AI News