In mid-September 2026, the U.S. National Science Foundation took a decisive step that signals a broader scientific realignment. As part of its $1.5 billion X-Labs initiative, NSF opened a formal solicitation for research proposals focused on Artificial Intelligence for Physical Systems. The agency also previewed two forthcoming topics, From Sequence to Function and Computation at the Limit of Physics, completing its 2026 portfolio of high-ambition research directions. This move is not merely another funding announcement. It marks a recognition that the most consequential advances in artificial intelligence are shifting from the generation of text, images and code toward systems that can model, control and discover behavior in the physical world.

For more than a decade, progress in digital AI has been measured largely by improvements in language models, diffusion systems and recommendation engines. Those capabilities transformed information work. Yet the physical domain robots that must grasp uncertain objects, scientific instruments that must tune themselves in real time, manufacturing lines that must adapt to material variability, and infrastructure networks that must respond to cascading failures remains far less tractable. NSF’s new X-Labs topic explicitly targets this gap. According to the agency’s September 16 announcement, “Decades of progress in digital AI have positioned the U.S. to redefine how intelligent systems interact with the physical world. The next frontier of AI requires developing intelligent, adaptive, and scalable systems for complex physical environments.” The solicitation seeks early-stage platform technologies in robotics, embodied systems, human-robot interfaces, advanced sensing and cyber-physical systems, with potential impact across healthcare, national defense, emergency response, advanced manufacturing and scientific discovery.

The X-Labs model itself is designed for this kind of challenge. Unlike traditional grants, X-Labs fund independent, full-time teams of researchers, engineers and entrepreneurs through Other Transaction agreements. Teams receive substantial multi-year resources, operational autonomy and milestone-driven funding that can scale to tens of millions of dollars annually once early concepts prove viable. Brian Stone, performing the duties of NSF Director, framed the approach clearly: “Science is moving faster than ever, and the challenges we face increasingly require new ways of working across disciplines and sectors. NSF X-Labs bring together researchers, engineers, and entrepreneurs to pursue ambitious scientific questions with the flexibility to test new ideas, learn quickly, and adapt along the way.”

This direction aligns with work already under way at national laboratories and research universities, where AI is moving from a post-processing analysis tool into an active experimental partner. At Pacific Northwest National Laboratory, the Center for Robotics and Autonomy has built ARCADIA Agentic Robotics and Curated AI Data for Intelligent Autonomy a platform that integrates AI agents, robotics and data systems into a single discovery ecosystem. The laboratory has committed more than $5 million annually through its Foundational Autonomy Investment to scale the approach. Bob Runkle, director of the Center, has stated the ambition without qualification: “At PNNL, autonomy will touch every experiment in five years, and the Center for Robotics and Autonomy will make that happen.” In practice this means AI agents that design experimental campaigns, translate goals into robot instructions, monitor live data streams, and adjust parameters such as temperature, concentration or flow rates while instruments are running. Early deployments already accelerate materials synthesis for battery research and microbial phenotyping, compressing design-build-test-learn cycles that once required weeks of manual effort.

Similar closed-loop systems are appearing at other Department of Energy facilities. At Lawrence Berkeley National Laboratory’s Advanced Light Source, researchers deployed the first language-model-driven agentic AI capable of autonomously executing multi-stage physics experiments on a production synchrotron. The system, known as the Accelerator Assistant, accepts natural-language requests from operators, retrieves historical archive data, resolves control-system channels, generates executable scripts, interacts with the accelerator under strict safety constraints, and returns analysis. In a representative machine-physics task, preparation time fell by two orders of magnitude compared with manual scripting by an expert, while operator-standard safety interlocks remained fully enforced. The architecture relies on plan-first orchestration, bounded tool access and dynamic capability selection design choices that keep the system auditable and reproducible. The results were published in Physical Review Research in early 2026, establishing a concrete blueprint for agentic AI in large-scale scientific infrastructure.

University-led efforts are reinforcing the same pattern. At the University of Chicago and Argonne National Laboratory, the Polymer Laboratory for AI, Robotics, Informatics and Standards (PoLARIS) received a $20 million NSF award to create a distributed cloud laboratory for soft-materials research. AI agents coordinate instruments across multiple buildings, running 100 to 300 experiments per day once fully operational. Researchers describe the system as “a distributed operating system for scientific discovery” in which digital twins allow workflows to be tested in simulation before physical execution. Parallel NSF-supported programmable cloud laboratory nodes, including the AURORA facility at the University of Utah, extend the model to quantum-device fabrication and cryogenic measurement, enabling remote researchers and AI agents to design, execute and analyze experiments over the internet.

These projects share several technical characteristics that distinguish them from generative AI systems. First, they operate under continuous physical constraints: energy limits, safety interlocks, material availability and real-time latency. Second, they must handle non-stationary environments sensor drift, instrument aging, batch-to-batch material variation that pure digital models rarely encounter. Third, they generate training data through interaction rather than consuming static corpora. The resulting datasets are often sparse, noisy and expensive, demanding new algorithms for sample-efficient learning, uncertainty quantification and safe exploration. Edge AI, distributed learning across robot swarms, and bio-inspired control architectures appear repeatedly in both the NSF solicitation language and the laboratory implementations.

The two forthcoming X-Labs topics extend the same logic deeper into the physical sciences. From Sequence to Function aims to close the remaining gap between predicted protein structure and controllable biological function. Platform technologies developed under this topic would enable intentional design of proteins or entire genomes whose behavior is predictable rather than merely discoverable. Computation at the Limit of Physics targets the energy-efficiency gap between conventional CMOS computing and fundamental physical bounds, seeking new paradigms that could sustain innovation over the next half-century. Both topics require the same experimental partnership between AI systems and physical apparatus that the AI-for-Physical-Systems solicitation prioritizes.

National laboratories occupy a distinctive position in this emerging ecosystem. Their large-scale user facilities accelerators, light sources, neutron sources and high-throughput chemistry platforms generate continuous high-dimensional data streams that are natural training grounds for adaptive AI. At the same time, the laboratories maintain rigorous safety cultures and multi-decade operational expertise that pure academic or startup teams often lack. The combination is producing systems that are neither purely autonomous nor purely human-operated, but hybrid: AI agents handle routine optimization and low-level control while human scientists set high-level objectives and intervene at decision points that carry scientific or safety risk.

Manufacturing and infrastructure applications are beginning to follow the same trajectory. Cyber-physical systems research supported by NSF’s longstanding CPS program now emphasizes neuro-symbolic and embodied AI approaches, safe learning under uncertainty, and co-design of hardware and software. Industrial partners are exploring edge AI for predictive maintenance of complex machinery and swarm coordination for distributed sensing networks. The common thread is that the AI system is no longer a passive analyzer of data collected by humans; it is an active participant that proposes interventions, executes them through actuators or robots, observes the consequences, and updates its internal model.

The technical barriers remain substantial. Physical systems are only partially observable. Control actions can have irreversible consequences. Simulation-to-reality gaps persist even with high-fidelity digital twins. Training data for rare failure modes is scarce by definition. Addressing these challenges requires advances in formal verification for hybrid systems, physics-informed neural architectures that respect conservation laws, and multi-agent coordination protocols that scale without central bottlenecks. The NSF X-Labs mechanism is deliberately structured to support the multi-year, multi-disciplinary teams needed to attack such problems systematically rather than incrementally.

What is emerging is less a single breakthrough technology than a new experimental paradigm. AI systems are becoming laboratory partners that can plan campaigns, operate instruments, interpret intermediate results and propose the next measurement all while remaining accountable to human oversight and physical constraints. The September 2026 NSF announcement formalizes this shift at the federal research level, connecting it to concurrent efforts at national laboratories and universities. The resulting body of demonstrated work closed-loop materials discovery platforms, agentic control of production accelerators, and distributed cloud laboratories provides a factual foundation for the claim that the next stage of AI progress will be measured not only by linguistic fluency but by the ability to interact productively with the physical world.

As these systems mature, the distinction between computational modeling and experimental practice continues to blur. An AI agent that can tune an accelerator lattice, synthesize a new polymer under closed-loop control, or design a protein with predictable catalytic behavior is no longer merely generating content. It is participating in the discovery process itself. NSF’s X-Labs investment, together with parallel programs at DOE laboratories and university consortia, is building the institutional and technical infrastructure required for that participation to become routine rather than exceptional.