Meet EcoBOT: The Autonomous Lab Standardizing Plant-Microbe Research

To illustrate how EcoBOT couples continuous measurement, adaptive modeling, and experimental redesign, the Berkeley Lab researchers used the system to observe how the model grass Brachypodium distachyon responds to environmental stressors such as nutrient deprivation and copper toxicity. In a traditional workflow, researchers might test a random spread of copper concentrations and wait weeks to measure the results. But inside EcoBOT’s compact cabinet, a robotic arm can autonomously manage over 150 individual EcoFABs simultaneously across three shelves. This robotic hardware doesn’t just automate the process; it intentionally maintains a highly controlled physical environment, providing the necessary foundation for the system-level modeling and downstream adaptive decision-making.

Historically, extracting continuous data from that many biological environments would have been a grueling, manual task prone to human error. To solve this, researchers equipped EcoBOT with a suite of Berkeley Lab-developed deep learning tools that serve as the system’s digital eyes. Addressing this sensing challenge required developing sophisticated new computer vision algorithms capable of reliably translating complex, noisy biological imagery into precise measurements.

Below ground, a tool called RhizoNet serves as an automated root tracker. Rather than relying on inconsistent manual interpretation of root images, RhizoNet uses neural-network-based segmentation to digitally separate fragile plant roots from the noisy background of the hydroponic fluid in a standardized and reproducible way. In validation tests, it successfully standardized the analysis of thousands of images, precisely tracking root growth dynamics across all the different copper treatments. Above ground, a computer vision tool called EcoSpec scans the plant’s shoots and analyzes complex, multi-wavelength hyperspectral images to monitor plant health. This tool has demonstrated high accuracy in high-throughput monitoring—while maintaining consistency across longitudinal measurements.

The EcoBOT becomes a true self-driving laboratory through the continuous interaction between its physical infrastructure, sensing systems, and adaptive modeling framework. The robotic hardware stabilizes the experimental environment, the imaging systems convert plant behavior into quantitative measurements, and gpCAM uses those measurements to identify where uncertainty is highest and determine which experiments should be performed next. Using Gaussian-process-based modeling, gpCAM analyzes preliminary experimental results, estimates uncertainty across the experimental landscape, and calculates the next copper concentrations that are likely to be most informative. By iteratively targeting these knowledge gaps, this autonomous approach improved the predictive accuracy of the plant biomass models by more than thirty percent. Training and processing the complex visual data for these advanced machine learning models requires massive computational power, which the team accesses using supercomputers at the National Energy Research Scientific Computing Center (NERSC).

“This level of automation now positions us to go after our long-term goal of using it to help elucidate beneficial plant-microbe interactions,” said Andeer. “I was actually collecting data while on the other side of the country, just by logging in and hitting ‘go.’ We no longer have to arrange for a team of research assistants to take individual measurements and hope they are recorded consistently. EcoBOT feeds those measurements directly into our models. And because it all happens inside the sterile EcoFAB environment, we can guarantee there are no outside microbes influencing the results, which is impossible in a greenhouse.”

Andeer notes that Berkeley Lab’s culture of team science was essential to realizing this vision. Bringing the self-driving lab to life required a collaboration of plant biologists, robotics engineers, and mathematicians from the Lab’s CAMERA team.

“We originally built gpCAM as open-source software because researchers at massive experimental facilities were simply drowning in data,” said Marcus Noack, a researcher in the Applied Mathematics and Computational Research (AMCR) Division and CAMERA, as well as the developer of gpCAM. “When you are exploring a vast, uncharted experimental landscape, measuring everything is impossible. Instead, gpCAM allows the instrument to calculate its own uncertainty and pinpoint the exact data points needed to complete the map. Whether you are scanning a 2D material or testing copper toxicity in an EcoFAB, the AI actively steers the experiment so we can learn as much as possible, as efficiently as possible.”