Between 2019 and 2023, three to five TerraSentia robots traversed fields to observe corn at scale, combining genetic data and climate conditions. They collected data from 198,249 plots across 142 sites, according to a study published in Communications Biology by DeBruin and coauthors. Human teams can perform these measurements, but time and labor constraints severely limit their deployment at scale and across multiple environments. This volume of data improves phenotyping efficiency and makes possible the analysis of interactions between varieties, nitrogen inputs, and environmental conditions—notably rainfall variations—at a scale inaccessible to manual methods.

The essentials

  • Between 2019 and 2023, three to five TerraSentia robots measured 198,249 corn plots across 142 sites, according to DeBruin et al., Communications Biology (2025).
  • This volume allows statistical distinction of interactions between varieties, nitrogen inputs, and environmental conditions, something that manual measurement makes severely limited and costly.
  • The bottleneck in variety selection is not genetics but field observation: robots tackle this link.
  • The study documents partial data openness and proprietary code, without establishing a binary opposition between public pooling and capture by integrated proprietary chains.
  • Signals to monitor are cost per plot, the share of freely reusable protocols, and the number of varieties tested under water stress or low-input conditions.

The bottleneck blocking variety selection

Ownership of these data and the agricultural direction they will produce remain indeterminate.

Corn genetics has progressed at a steady pace since the 1990s. Sequencing tools made reading the genome quick and inexpensive. But selection does not stop at the genome: it requires knowing how a plant behaves in a real field, under real rain, with a given dose of nitrogen. This is phenotyping, and it remained slow, expensive, and human-dependent.

A breeder wanting to test a thousand varieties in twenty different environments must observe twenty thousand plots. With field teams, this takes entire seasons and budgets that only large seed companies can mobilize. Public institutes and cooperatives abandon hypotheses for lack of observation capacity, not lack of ideas. This is the bottleneck that DeBruin and coauthors’ study quantifies, demonstrating at this scale the feasibility of reducing the phenotyping bottleneck through robotic deployment.

TerraSentia robots measure or estimate plant height, leaf area index, stem diameter, and canopy architecture; these traits can serve to estimate biomass, without constituting a direct measurement of aerial biomass. They move autonomously between rows, capture multisensor data, and produce comparable data series from one site to another. For 4,826 units, the authors estimate roughly 20 hours of robotic data collection against 150 hours of manual measurement of ear height. Robotics greatly reduces human labor time for certain measurements and is presented as a low-cost approach, lowering the entry barrier to more ambitious experiments.

Data from 198,249 plots in service of observation

Volume is not an end in itself. The large number of plots and environments makes possible the analysis of interactions that require large-scale phenotypic collection.

Agronomists have long sought to understand interactions between a plant’s genotype, the dose of nitrogen it receives, and the site’s climate conditions. These interactions, noted GxExM in the literature, lie at the heart of selection for reduced-input agriculture. One variety can excel with abundant nitrogen and collapse under stress; another can hold up under drought but disappoint in favorable years. Distinguishing these profiles requires testing them across many environments simultaneously.

With 142 sites and 418 growing seasons, the study published in Communications Biology provides exactly this comparative context. Multisensor data allow linking morphological traits measured to final yields, and testing whether this relationship is stable or variable across environments. Traits like leaf area index or panicle height become predictors of final yield, usable well before harvest to guide crosses.

The robot generates data; breeders decide. The machine is an observation instrument, not an agronomic decision-maker. This roboticized collection can make phenotyping denser and less demanding in labor, with potential to accelerate selection.

The economics of research, before the plant itself

DeBruin and coauthors’ result interests research economists as much as agronomists. It documents that field phenotypic observation constitutes a major bottleneck in maize selection and research. Resolving this bottleneck through robotics could alter access to field observation and associated costs for different research actors.

A public university in the American Midwest, a Brazilian agricultural institute, or a French grain cooperative could, in theory, deploy a few robots across their trial networks and collect data comparable to that of major seed firms. The investment remains significant, but it is far more accessible than building sequencing capacity or a genomics platform. Making field observation accessible is possible, provided that protocols and data formats are standardized and open.

It is precisely here that institutional choices play a decisive role. Data from the DeBruin et al. study, including trials involving Corteva, are deposited in full on Dryad; EarthSense’s processing code remains proprietary. By contrast, data collected by Bayer or other actors in their proprietary trials are subject to access and use restrictions, while certain summaries and trial results are published. The models they feed, the varieties they help select, the intellectual property rights that result: all of this remains within the firm’s value chain. Robotics accelerates research, but it accelerates according to the priorities of those who own the robots and the data.

This debate is not unique to agronomy. The question of who controls the infrastructure for producing knowledge also runs through basic research. French research publishes more internationally but weighs less on standards: a signal that the capacity to produce results is insufficient if others set the standards and capture the uses.

Nitrogen sobriety and resilience: agronomic objectives long underfunded

The nitrogen crisis is one of agriculture’s costliest problems, yet one of the least visible in public debate. Synthetic nitrogen enabled feeding billions of additional people in the twentieth century. It today consumes roughly 1 to 2 percent of global energy for its production, and its excess in soils and waters causes massive ecological damage, from coastal dead zones to groundwater pollution.

Selecting corn varieties capable of producing adequately with less nitrogen is therefore an agronomic and environmental objective of the first order. It is also a commercially unattractive objective for major seed companies, whose flagship varieties sell precisely in intensive systems where nitrogen is abundant. Low-input trials have long been underrepresented in breeding programs, for lack of observation capacity at the necessary scale to detect subtle varietal differences.

The 418 growing seasons in the study include varied input conditions. This approach can help study varietal responses to nitrogen and contrasting environments, including rainfall differences. Public breeders, cooperatives, and institutes specializing in low-input agriculture would gain an advantage if this observation infrastructure became accessible to all.

The dataset underlying the study was deposited on Dryad, the standard open data platform for life sciences. This is a positive signal, but partial: data from a single experiment are insufficient to constitute a shared phenotyping platform. Common protocols, interoperable robots, and sustainable network operation over time remain to be built.

Who will own the infrastructure for observing life

The 2030–2040 horizon outlines two plausible trajectories, and they do not lead to the same place.

In the first, shared phenotyping networks emerge among universities, public institutes, and cooperatives. Robots are rented or pooled, measurement protocols are standardized, data are deposited in open and reusable databases. The cost per measured plot continues to fall, and breeders working on so-called neglected crops—sorghum, millet, cowpea, varieties suited to African and Asian smallholder agriculture—can access the same type of infrastructure as major firms. This trajectory requires stable public support and political will to condition investment on open data standards.

In the second, major seed actors integrate robots, sensors, machine learning models, and variety rights into closed proprietary platforms. They innovate faster, and some of their innovations genuinely benefit farmers. But the directions of selection reflect their priority markets: grain corn in favorable conditions, maximum yield, compatibility with their crop protection lines. Stress-adapted varieties and secondary crops remain off the radar, for lack of data concerning them. Agronomic knowledge accumulates, but it concentrates.

Signals that would allow distinguishing these trajectories are observable. Cost per plot and per measured trait will indicate the speed of access for non-commercial actors. The share of phenotyping protocols deposited in free access, and data actually reusable under open license, will say whether the public scientific community can build on these results or must reproduce them at its own expense. The number of varieties tested under water stress or low nitrogen in major trial networks will indicate whether research follows global food priorities or solvent markets.

These trade-offs will not be made spontaneously in favor of openness. Directing shared phenotyping platforms among universities, public institutes, and cooperatives, and conditioning public agricultural research support to reusable data standards, are political decisions. They resemble debates around health data or geographic data, where the analogy with digital infrastructure becomes obvious. The book review “Radical Infrastructure” recalls that choices in technical architecture are never neutral: they predict who will be able to use a system and for what purposes.

Potential contributions for African and Asian breeders

The question of access to phenotyping infrastructure goes beyond debates between American firms and Western public institutes. The food crops that feed the majority of the world’s population, in low-input agricultural systems, with rising water stress, are precisely those for which phenotyping data are rarest and most necessary.

The Consultative Group on International Agricultural Research, the global network of international agricultural centers, has for decades run variety breeding programs for crops like irrigated rice, wheat, tropical corn, and cassava. These programs work with constrained human resources. Access to phenotyping robots at affordable costs, and to data produced according to protocols compatible with theirs, could substantially accelerate their work. But this access is not automatic.

The DIGITAfrica initiative documents how scientific communities can build shared data infrastructure to reduce their dependence on external platforms. Roboticized crop phenotyping follows comparable logic: benefits will be more widely distributed to the extent that infrastructure is designed from the outset as a shared good, with open standards and affordable access.

The trajectory of agronomic innovation for coming decades is being determined partly now, in governance choices made by universities, research agencies, and governments around the data platforms being built. DeBruin and coauthors’ study shows what becomes technically possible. It does not say who will benefit.

Sources

  1. DeBruin et al., “Breaking the field phenotyping bottleneck in maize with autonomous robots”, Communications Biology, 2025. https://www.nature.com/articles/s42003-025-07890-7
  2. Associated dataset, Dryad. https://datadryad.org/dataset/doi%3A10.5061/dryad.m905qfvb6