DOE Said 2-3 Awards and Made 4. The Robotics Testbed Arithmetic Says Every Winner Got Cut.

Two numbers from the same solicitation, five months apart, that cannot both be true at face value.

In LAB 26-3601, issued May 14, 2026, the Department of Energy's Office of Science stated its expectations plainly: 2-3 awards, at $3,000,000-$5,000,000 per year each, over a 3-year project period, against a total of $30 million in current and future-year funds — of which $2 million is available in FY26 and $28 million in the outyears.

On October 8, 2026, at the Golden Age of Science Summit in Washington, DOE announced four selections.

Run the arithmetic. Four awards sharing $30 million over three years is $7.5 million per award total, or $2.5 million per award per year. The announcement's stated floor was $3 million per year. Every winner is below the advertised minimum, and the median winner is roughly half what the solicitation's midpoint implied.

That is not a scandal. It is the single most useful piece of information in the release, because it tells you how DOE resolves oversubscription in the Genesis Mission era: it widens the field and thins the portions.

What got funded

Four national laboratories lead four testbeds, each attacking a different failure mode of autonomous laboratory science.

MAESTRO — Modular Autonomous Experimentation through Self-improving Testbeds for Robotic Operations, led by Argonne National Laboratory. Modular, self-improving infrastructure for laboratory robotics, combining digital twins and world models, reusable robot skills, agent-based orchestration, safety mechanisms, and learning from operational experience, with evaluation of capability improvement and transfer across instruments and environments.

DART — DOE AI Robotics Testbed to Generalize Autonomous Science, led by Brookhaven National Laboratory. DART's question is the sharpest of the four: when is robotic autonomy reliable, transferable, and scientifically valid? The project builds adversarial digital counterparts, a verified failure atlas, multimodal robot-learning methods, and approaches for transferring task knowledge across robotic platforms — evaluating not merely whether a robot completed a motion, but whether the resulting physical and scientific outcomes meet the requirements of the experiment.

TRACE — Testbed for Robotics and Autonomy in Connected Experiments, led by Oak Ridge National Laboratory. Builds on Oak Ridge's in-house INTERSECT automated laboratories ecosystem to connect robots, scientific instruments, AI agents, digital twins, data systems, and computing resources through standardized interfaces, with emphasis on testing and deployment across laboratories and scientific domains.

SPIRE — A Source-to-Discovery Platform for Instrumentation, Robotics, and Embodied AI in DOE Photon-Science Facilities, led by SLAC National Accelerator Laboratory. Integrates physical sample manipulation, intelligent detectors, accelerator and instrument controls, persistent experiment state, and edge-to-HPC computing to demonstrate source-to-discovery autonomy at photon and electron-science facilities. SPIRE's named partners include Stanford University, the University of Chicago, Argonne, Brookhaven, and Lawrence Berkeley National Laboratory, with Stanford faculty Karen Liu, Monroe Kennedy III, Eric Darve, and Mert Pilanci involved. SPIRE lead PI Angelo Dragone framed the stakes: "SI-assisted automation is the future of research in science and technology, and SLAC is well-positioned to lead this momentous step."

The SPIRE problem statement is the clearest illustration of why DOE is spending here at all. LCLS-II generates data at a rate exceeding one terabyte per second. No human team, on any shift arrangement, analyzes a stream that wide in real time. The instrument has outrun its operators.

The eligibility wall, and the two doors in it

LAB 26-3601 is categorical: "This is a DOE National Laboratory-only Announcement. FFRDCs from other Federal agencies are not eligible to submit in response to this Program Announcement."

No university leads. No company leads. No other-agency FFRDC leads. Each DOE/NNSA national laboratory is limited to leading one proposal, with no limit on the number of proposals where an institution appears as a subrecipient. An individual may not be PI on more than one proposal, though a PI may appear as senior or key personnel — including on subawards — on unlimited other submissions.

PIs must hold permanent positions at the applicant institution, whether tenured, tenure-track, or a staff appointment. Individuals in term-limited appointments — adjunct, visiting faculty, fellows, or similar — are not eligible to be proposed as PI. Joint appointees qualify only if the work is performed at the applicant institution and the PI is a paid employee of it. Someone paid by another institution may not be named PI, but may be named in other senior or key roles.

Cost sharing is not required.

And then, inside a labs-only announcement, two mandates that exist specifically to let everyone else in:

External Collaborator Required: Each proposed testbed thrust must include at least one external collaborator from outside the DOE laboratory complex.

Industry Participation Required: Each testbed thrust must include specific industry participation, either as funded external collaborators or in advisory roles.

Those two sentences are the entire opportunity surface for anyone not employed by a national lab. Read them with the proposal structure, which caps each laboratory's portfolio at at most three distinct testbed thrusts, each led by a Laboratory Senior/Key Person.

Three thrusts per lab. Each thrust requires at least one non-lab external collaborator and substantive industry participation. Four winning labs, therefore, carry up to twelve thrusts, each with mandatory non-lab seats — as many as 24 external collaborator and industry positions across the four projects, structurally required by the announcement.

That is the real answer to "how do I get into a labs-only program." You do not. You get into a thrust.

What DOE wants from each kind of outsider

The announcement is specific about the value an external collaborator is expected to bring, and it is not additional robotics expertise. External collaborators are expected to collectively bring "perspective on needs across user communities and scientific domains, and experience with broadly adopted platforms and interfaces," plus "expertise helpful for accelerating maturation, interoperability, and dissemination of successful testbed capabilities (including documentation, benchmarking practices, and community adoption pathways)."

Translate that into who gets picked. The external collaborator seat goes to whoever can credibly say: I represent a user community that will adopt this, and I know which interfaces they already use. That is a maintainer of a widely used instrument-control or lab-automation framework, a benchmark steward, a facility user-program lead, a standards-body participant — or a domain scientist who runs the kind of campaign the testbed is meant to serve.

It does not go to the group with the best manipulation paper. DOE is buying adoption pathways, not capability.

For industry, the bar is different and stated with an unusual guardrail. ASCR "expresses no general preference for whether industry participants are funded or cover their own costs." Participation must be substantive and clearly tied to testbed objectives — "hardware/software integration, interface standardization, deployability considerations, or technology transition planning." And then the line that kills the easy pitch: "Work proposed must not be a supplement to existing industry investment, but should enable new capabilities and testbed-driven evaluation on the scientific frontier."

A robotics vendor cannot get its existing product roadmap funded by attaching it to a thrust. The work must be additive to what the company was already doing, and it must be evaluated on the testbed.

Six ways to be declared out of scope

The announcement's out-of-scope list functions as the responsiveness screen, and most of its entries describe work that is individually excellent:

• Proposals primarily focused on a single scientific application without broader testbed reuse value.
• Efforts that do not include a substantive robotics/automation testbed component.
• Projects restricted to biomedical, agricultural, warehouse, or consumer robotics use cases that do not translate to DOE laboratory science missions.
• Efforts whose primary goal is incremental hardware refresh or commercial deployment without a testbed-driven capability and evaluation objective.
• Software-only AI/ML efforts not evaluated on, or co-designed with, physical robotic systems in representative laboratory workflows.
• Pure teleoperation interfaces without a pathway to training, evaluation, and progression toward scalable embodied autonomy within the testbed.
• Efforts emphasizing cybersecurity, networking, or facility management without an experiment-centric robotics/autonomy testbed objective.

Note how many of those exclusions target success. A warehouse robotics team with a working stack is out. A foundation-model group with state-of-the-art benchmarks and no physical system is out. A vendor with a mature product is out. The announcement states the governing principle directly: "Testbeds are not intended to be one-off demonstrations. Rather, they should provide reusable infrastructure that enables the development and evaluation of multiple workflows and can be extended over time."

The capability you have built is the thing most likely to disqualify you, if it is not reusable by someone else.

The metrics are the specification

Buried in the program description is the most actionable content in the document — the quantitative measures proposals are expected to commit to:

• time-to-result
• mean time between human interventions
• reproducibility and provenance
• task success rate
• generalization across tasks or sites

"Mean time between human interventions" deserves attention. It is a borrowed reliability metric, and it reframes the entire problem. A demo that works once with a graduate student standing beside it scores near zero. A clumsier system that runs four days unattended scores high. Proposals are "expected to culminate in one or more integrated demonstrations that exercise the testbed in a realistic laboratory or high-fidelity representative setting" with a clear plan for quantitative performance measurement.

The underlying technical concept DOE names is embodied operational intelligence (EOI): the ability for robots to acquire and refine laboratory skills through structured task design, human demonstration and supervisory operation, and data-driven adaptation. The announcement offers one illustrative approach — state-of-the-art foundation models, "e.g., vision-language-action or related models," adapted to laboratory tasks through few-shot learning, imitation learning, or other data-efficient methods — while explicitly leaving teams free to choose their own.

The long-term vision is stated as an "instrument of instruments" paradigm, in which automated workflows couple experiment execution, data analysis, and adaptive refinement into iterative scientific campaigns.

Responsive proposals may address modular and interoperable testbed architectures with standardized interfaces and APIs; digital-twin and simulation-enabled testbeds supporting task rehearsal and constraint tuning; multimodal perception and time-synchronized data pipelines; safety, compliance, and hazard-emulation capabilities including safety envelopes, interlocks, and fail-safe behaviors under radiological, chemical, thermal, or mechanical constraints; benchmarking and standards development; and operator and researcher training.

That hazard-emulation line explains the entire eligibility structure. The experimental environments where autonomy pays off most are the ones humans cannot safely occupy — and those environments exist at national laboratories.

Reading the $2 million

The FY26 money is $2 million. Split four ways, that is roughly $500,000 per project in the current fiscal year. The remaining $28 million is explicitly "subject to the availability of future year appropriations."

That shape has consequences worth planning around:

Year one is a staffing year, not a buildout year. Half a million dollars funds a small team writing architecture and integration plans. Capital equipment purchases largely wait for outyear money.

Outyear risk is real and distributed. Ninety-three percent of the program's value sits in appropriations that have not happened. A thrust-level external collaborator negotiating a subaward should understand that the committed money is thin and the rest is a forecast.

Widening from three awards to four was likely deliberate under this constraint. When the current-year appropriation is $2 million regardless of how many projects you fund, adding a fourth project costs almost nothing today and buys an additional lab-complex capability, an additional set of external collaborators, and an additional constituency for the outyear request. The cost of the decision lands in FY27 and beyond.

This is the same pattern visible across DOE's FY26 Office of Science portfolio. The Early Career Research Program advertised roughly $145 million and delivered $96 million. The companion LAB 26-3603 quantum validation and verification testbed carries $45 million with $14 million in FY26 — a far healthier first-year ratio. Advertised program totals and realized awards are diverging, and the direction is consistent.

Where this fits in the Genesis Mission

LAB 26-3601 explicitly requires deliverables that "generalize across workflows and sites, supporting broad adoption across DOE laboratory environments and the Genesis Mission platform."

That one clause is the structural point. These four testbeds are not independent robotics projects. They are infrastructure for a larger program — the same program behind DOE's 26 AI grand challenges at $320 million, the $159 million in Phase II Genesis awards, and the $100 million Genesis Mission graduate fellowship pilot.

Genesis-adjacent awards also flow through other doors. UT Austin, for instance, reports five teams funded under Genesis Mission lines, including Professor Volkan Isler's SAFE-BOLT project — Safety-Assured Force-Aware Execution for Bimanual Operations on Lab Tools — developing force-aware robotic assistants to automate labor-intensive experiments, in collaboration with the Materials Discovery Research Institute and the robotics startup Medra AI. That is a university-led award in the same technical space as LAB 26-3601, won through a different mechanism. The lesson: when a labs-only door closes, check whether the same capability is being bought somewhere else in the program.

The expected deliverable set across the four testbeds is unusually public-goods oriented: open software and interfaces, reusable robot skills, digital-twin environments, benchmark tasks, datasets and trained models, safety practices, provenance records, and training resources.

If that list materializes, the most valuable outputs of a $30 million program will be free. Benchmark tasks, datasets, and standardized interfaces are usable by any lab anywhere — and a group that builds against those benchmarks early is positioned for whatever the next solicitation turns out to be.

What to actually do

If you are at a DOE national laboratory: the lead-proposal lane for this cycle is closed. The live question is whether your capability belongs inside one of the four funded portfolios as a thrust participant or subrecipient — remembering that there is no limit on subrecipient appearances, and that a PI elsewhere can be senior personnel on unlimited other submissions.

If you are at a university: stop looking for a solicitation and start looking for a thrust lead. Up to twelve thrusts across four projects each require an external collaborator. Approach the named PIs with the thing DOE said it wanted — a user community, a broadly adopted interface, a benchmark practice, an adoption pathway — not with your best robotics result. And watch for the university-eligible adjacent lines; SAFE-BOLT demonstrates they exist.

If you are in industry: every thrust needs you, and funded participation is explicitly on the table alongside advisory roles. Prepare the answer to the hard question before the first call: what new capability does this enable that is not a supplement to what you were already building? Interface standardization and deployability are named as acceptable contributions. Product roadmap acceleration is not.

If you are planning for the next cycle: build to the metrics. Time-to-result, mean time between human interventions, reproducibility and provenance, task success rate, and generalization across tasks or sites are now the vocabulary of this portfolio. A lab that instruments its own workflows against those five numbers today will have evidence nobody else has when the next announcement issues — and in a program that pays for demonstrated throughput rather than described plans, evidence is the entire proposal.