Open Access Article
This Open Access Article is licensed under a
Creative Commons Attribution 3.0 Unported Licence

Automating care by self-maintainability for full laboratory automation

Koji Ochiai a, Yuya Tahara-Arai bcd, Akari Kato ade, Kazunari Kaizu af, Hirokazu Kariyazaki g, Makoto Umeno h, Koichi Takahashi *ai, Genki N. Kanda *adj and Haruka Ozaki *bdk
aLaboratory for Biologically Inspired Computing, RIKEN Center for Biosystems Dynamics Research, 6-7-1 Minatojima Minamimachi, Chuo-ku, Kobe, Hyogo 650-0047, Japan. E-mail: ktakahashi@riken.jp
bBioinformatics Laboratory, Institute of Medicine, University of Tsukuba, 1-1-1 Tennodai, Tsukuba, Ibaraki 305-8577, Japan
cPhD Program in Humanics, School of Integrative and Global Majors, University of Tsukuba, 1-1-1 Tennodai, Tsukuba, Ibaraki 305-8577, Japan
dLaboratory Automation Suppliers' Association, 2-1-8 Minatojima Minamimachi, Chuo-ku, Kobe, Hyogo 650-0047, Japan
eResearch DX Foundation Team, RIKEN Data and Computational Sciences Integration Research Program, 6-7-1 Minatojima Minamimachi, Chuo-ku, Kobe, Hyogo 650-0047, Japan
fCell Modeling and Simulation Group, Exploratory Research Center on Life and Living Systems, National Institutes of Natural Sciences, 5-1 Higashiyama, Myodaiji, Okazaki, Aichi 444-8787, Japan
gRobotics Division, YASKAWA Electric Corporation, 2-1 Kurosaki Shiroishi, Yahatanishi-ku, Kitakyushu, Fukuoka 806-0004, Japan
hBAIKEIDO LLC, 3-8-1 Asano, Kokurakita-ku, Kitakyushu, Fukuoka 802-0001, Japan
iGraduate School of Media and Governance, Keio University, 5322 Endo, Fujisawa-shi, Kanagawa 252-0882, Japan
jDepartment of Robotic Science, Medical Research Laboratory, Institute of Integrated Research, Institute of Science Tokyo, 1-5-45 Yushima, Bunkyo-ku, Tokyo 113-8519, Japan. E-mail: genki.kanda@tmd.ac.jp
kLaboratory for AI Biology, RIKEN Center for Biosystems Dynamics Research, 6-7-1 Minatojima Minamimachi, Chuo-ku, Kobe, Hyogo 650-0047, Japan. E-mail: haruka.ozaki@riken.jp

Received 15th April 2025 , Accepted 17th June 2025

First published on 19th August 2025


Abstract

The automation of experiments in life sciences and chemistry has significantly advanced with the development of various instruments and artificial intelligence (AI) technologies. However, achieving full laboratory automation, where experiments conceived by scientists are seamlessly executed in automated laboratories, remains a challenge. We identify the lack of automation in planning and operational tasks—critical human-managed processes collectively termed “care”—as a major barrier. Automating care is the key enabler for full laboratory automation. To address this, we propose the concept of self-maintainability (SeM): the ability of a laboratory system to autonomously adapt to internal and external disturbances, maintaining operational readiness. This ability is inspired by the homeostasis, resilience, autonomous state recognition, and adaptability seen in living cells. A SeM-enabled laboratory features autonomous recognition of its state, dynamic resource and information management, and adaptive responses to unexpected conditions. This shifts the planning and execution of experimental workflows, including scheduling and reagent allocation, from humans to the system. We present a conceptual framework for implementing SeM-enabled laboratories, comprising three modules—Requirement manager, Labware manager, and Device manager—and a Central manager. A laboratory design that is aware of SeM not only enables scientists to execute envisioned experiments seamlessly but also provides developers with a design concept that drives the technological innovations needed for full automation.


1 Automating care for full laboratory automation

Science is an activity to extend the wisdom of humankind, and the automation of science is the ultimate way to accelerate this activity.1–3 In the fields of life sciences and chemistry, significant advancements have been made in automating experimental operations, exemplified by the widespread use of automated pipetting machines and laboratory robots.4–13 However, automating experimental operations alone is insufficient to fully realize the potential of automated scientific discovery. Research inherently involves iterative processes, such as optimizing reaction conditions or exploring parameter spaces, which require adaptive and continuous decision-making. To address this need, the concept of closed-loop experimentation, or self-driving labs (SDLs), have emerged,14 which combines automated experiments with data-driven decision-making powered by artificial intelligence (AI). These integrated systems have demonstrated remarkable success in various fields, including photocatalysis, microbial genetics, and regenerative medicine.15–20 Research to increase the autonomy of these automated laboratories includes: error handling through reactive workflows,21 collaboration in the same space as humans using mobile robots,22 experiment execution while interacting with the experimenter,23 conversion of literature written in natural language into standard code that can be executed automatically,24 dynamic resource allocation,21,25 goal-oriented programming and instruction writing,26,27 fully closed-loop operation.28

While the integration of robotic automated experiments and AI has enabled the automation of iterative optimization processes, these systems still rely heavily on human involvement for what we propose to define as “care.” Care refers to the human-managed tasks required to maintain and support automated systems, encompassing a range of essential responsibilities at both the planning and operational steps of automated laboratories (Fig. 1). In the planning step, care includes translating experimental objectives into machine-readable workflows and scheduling resources such as equipment and reagents. In the operational step, it involves preparing and cleaning up experiments, restocking supplies, monitoring system states, and addressing errors or unexpected conditions. For example, humans must ensure compatibility between workflows and equipment limitations, such as splitting a 200 μL aliquot into two 100 μL portions to align with equipment constraints. Additionally, tasks like verifying equipment availability and adapting workflows to real-world constraints remain outside the scope of current automation capabilities, leaving these responsibilities as a bottleneck in achieving fully autonomous scientific discovery.


image file: d5dd00151j-f1.tif
Fig. 1 Human efforts in a conventional laboratory automation workflow (as is). This schematic illustrates how human specialists and automated systems collaborate to carry out experiments in biology or chemistry. Bold letters represent human efforts (care) in conventional laboratory automation. The research and development (R&D) team (blue area)—comprising domain specialists (e.g., biologists, chemists), hardware engineers, software engineers, and operators—collectively designs, plans, and executes experiments, creates AI-compatible workflows, and develops robotic procedures. The automated laboratory (pink area) is controlled by artificial intelligence (AI)/Software and houses robots and other automated machines responsible for executing tasks such as sample handling and manipulation. However, human operators remain essential for quality assurance and troubleshooting—for instance, addressing issues like pipetting errors or bubbles—along with setting up and cleaning the lab environment. The human laboratory and storage (green area) require manual tasks such as refilling consumables and preparing specialized equipment. In many cases, data collection also relies on human intervention, depending on the type of observation or instrumentation. While automation can streamline experiments, human care and oversight are critical for ensuring accuracy, maintaining equipment, and resolving unexpected problems. The instruments shown in this figure are examples and represent only part of the available devices: Maholo LabDroid (Robotic Biology Institute Inc.), Fluent (TECAN), Chemspeed (Chemspeed Technologies Inc.), OT-2 (Opentrons Labworks, Inc.), Cobotta Pro (Denso Wave Incorporated), BioStudio-T (Nikon Corporation).

Care is indispensable for making experimental protocols feasible under laboratory conditions, bridging the gap between conceptual designs and their practical execution. Usually, experimental protocols are abstract and must be adapted to the physical constraints and conditions of the laboratory: Even with the same conceptual protocol, the operational protocol can differ based on the laboratory environment and setup, due to variations in equipment, space, or resources. Such adjustment processes requires intervention of human managers and operators, primarily because (1) most of the conventional laboratory automation systems lack the ability to recognize their environments and adapt to them, (2) these systems are designed to operate within predefined constraints, relying on user-provided inputs such as initial setups and experimental conditions, (3) the dynamic nature of laboratory environments, where conditions can shift due to concurrent experiments or resource usage, necessitates manual intervention to prepare the setup immediately before execution.

This reliance on human care constrains the capabilities of automated laboratories, limiting their application primarily to repetitive or screening-based experiments including closed-loop experimentation. In the planning step, the dependency on human care restricts the range and complexity of experiments that can be executed. For example, when introducing new experiments, reagents, and labware, robotic protocols are iteratively refined through trial and error until human operators ensure that workflows are feasible, avoiding flexible modifications in experimental conditions, addition of newly performed experimental protocols, and parallel execution of multiple experiments. In the operational step, the dependency on human care (e.g. handling errors) limits the number of experiments that can be executed simultaneously, creating a bottleneck in throughput. As a result, the full potential of automation remains unrealized, with current systems unable to perform the adaptive and flexible processes required for more diverse and complex scientific discovery.

In an ideal automated laboratory, experiments are conducted entirely without human intervention. To achieve this, not only experimental operations but also the tasks we define as care must be fully automated. This includes addressing dynamic and unforeseen changes in the laboratory environment, ensuring that experimental processes can continue progressing even in the face of unexpected situations. Such a laboratory would require a new design paradigm in which the system itself autonomously acquires information, makes adaptive decisions, and manages its environment. This approach moves beyond the current reliance on humans to oversee and validate each step, paving the way for a more flexible and truly autonomous scientific discovery process.

In this vision, the entire laboratory functions as a cohesive system, akin to a single, integrated workstation. Instead of viewing individual devices as isolated units, the laboratory operates as a unified whole. Equipment management, labware tracking, waste disposal, and overall status monitoring are automated and coordinated as part of the system. By transferring these responsibilities—conventionally managed by humans—to the automated laboratory itself, the burden on human operators is eliminated. This enables the system to conduct not only repetitive or screening-based experiments but also highly complex, non-standardized workflows that require adaptability and creativity, fulfilling the broader potential of scientific automation.

Implementing automations of care through bottom–up extensional engineering is presumably difficult because there are no limits to what can be done. For instance, if there are multiple locations where a tube containing ethanol could be placed, a mechanism could be created to check each of them. However, such a process would be specific to a particular experiment and would require adjustment and redevelopment each time the type of experiment changes. This approach lacks scalability and adaptability for dynamic laboratory conditions.

2 Self-maintainability (SeM) for automating care

Here, to address these challenges, we propose a new concept for ideal automated laboratories called self-maintainability (SeM):

Self-maintainability (SeM) – the system's ability to maintain itself to the state where it can function effectively in environments subject to material consumption, equipment wear, and potential operational disruptions caused by external disturbances.

Conventional approaches to laboratory automation have primarily focused on repeating a single protocol multiple times (Fig. 2A). However, an ideal automated laboratory should be capable of automatically conducting a wide variety of daily experiments for multiple users (Fig. 2B). Such an automated laboratory can be characterized by three essential requirements. First, it must integrate multiple devices that operate in concert (Fig. 2C). Second, to accommodate diverse demands, the laboratory must flexibly incorporate and remove various elements including samples, labware, reagents, waste, and human operators (Fig. 2D). Third, it must function amid uncontrollable changes, such as ongoing experimental processes, environmental fluctuations, and human intervention (Fig. 2E). Historically, humans have provided the care required to address these challenges, but this care can be automated by incorporating processes analogous to those in living cells—specifically, “intake,” “discharge,” and “metabolism.” Consequently, the laboratory must be able to maintain its own function autonomously: regardless of resource consumption by ongoing experiments, sudden environmental changes, or unforeseen interruptions, it should continue to accept and execute experiments. This ability is here defined as SeM (Fig. 2F).


image file: d5dd00151j-f2.tif
Fig. 2 Conceptual overview of an ideal automated laboratory (A) once a human has completed the initial setup, the experimental robot or workstation can execute a single protocol without human intervention; after the run is finished, a human must return to clean up and re-set the system before the next experiment. Labware, samples, and reagents are processed automatically, producing waste and results. The gray rectangle represents the experimental robot or workstation. (B–F) (B) and (F) present the whole picture of an ideal automated laboratory. (C)–(E) show its key components. (B) The gray circle denotes the automated laboratory in its ideal state, where multiple orders (X, Y, Z) trigger automated experiments that generate corresponding results. (C) The automated laboratory relies on various components—robots, transporters, machines, and humans—that must cooperate to fulfill experimental demands. (D) In conducting experiments, the laboratory must intake necessary materials and discharge waste, processes that can generally be predicted in advance. (E) The laboratory is subject to various internal and external disturbances, including human intervention. These disturbances cause the situation to change moment by moment, posing obstacles to achieving the ideal state shown in (B). (F) Our newly proposed concept, self-maintainability (SeM), integrates state recognition and adaptive response. This allows the automated laboratory to maintain its function despite internal and external disturbances, echoing—only by analogy—the homeostasis and resilience of living cells or organisms.

The idea of SeM can be used to provide a holistic framework for designing systems capable of autonomously adapting to and mitigating real-world challenges (Fig. 3). Existing experiment automation, including most closed-loop experiment systems, has almost no SeM and relies on human care, whereas automated laboratories with high SeM can be operated with low burden on humans. In other words, today's automation requires users to think about things that are not important to them (e.g., where to place plates and tube racks on the desk, where to store consumables) each time an experiment is performed. Laboratory's SeM improvements will enable automation of experiments without the need for human effort or time for such matters. For example, tasks currently performed during the planning step, such as translating experimental objectives into machine-readable workflows and scheduling equipment and reagent use, would be autonomously managed by the system. Similarly, operational step tasks like preparing and cleaning up experiments, restocking supplies, monitoring system states, and resolving errors would be handled automatically. SeM not only eliminates repetitive human manual interventions but also enables laboratories to execute more complex and flexible experiments, overcoming the current bottlenecks in scalability of experiment type and adaptability. Furthermore, SeM can bridge the gap in expertise among human scientists, enabling researchers from different fields to conduct experiments effectively without requiring specialized operational knowledge. For instance, when a chemist seeks to test a hypothesis in another field such as biology, they may lack the tacit laboratory knowledge required to specify each cell-culture step or DNA-analysis parameter; an SeM system can translate that high-level objective into domain-appropriate, concrete procedures.


image file: d5dd00151j-f3.tif
Fig. 3 Self-maintainability (SeM). We define self-maintainability (SeM) as the system's ability to maintain itself to the state where it can function effectively in environments subject to material consumption, equipment wear, and potential operational disruptions caused by external disturbances. In the “without self-maintainability” scenario (top), a full R&D team must perform care and maintenance tasks (e.g., refilling the medium, verifying placement, teaching, and troubleshooting). By contrast, in the “with self-maintainability” scenario (bottom), a single scientist issues high-level instructions—such as “replace medium at 80% cell density”—while the system autonomously handles its own care.

3 SeM-enabled laboratories: redefining the roles of humans in laboratory automation

Hereafter, an automated laboratory with SeM will be referred to as a SeM-enabled laboratory. Conventional laboratory automation (including workflow management system, liquid handler, etc.) and SeM-enabled laboratories differ significantly in terms of their goal, scope, user assumptions, and handling of information (Fig. 4). The goal of conventional laboratory automation is to accurately follow the user's instructions, whereas SeM-enabled laboratories aim to fully realize the user's intentions. In terms of scope, conventional systems are limited to single or multiple robots, while SeM-enabled laboratories view the entire laboratory environment as a single, integrated system. Regarding user assumptions, conventional systems treat users as ideal or perfect entities who always provide clear and precise instructions. In contrast, SeM-enabled laboratories acknowledge that users' instructions are inherently incomplete and treat users as sources of uncertainty. In handling information, conventional automation assumes all necessary information is complete and that conditions can be fully calculated or confirmed in advance. On the other hand, SeM-enabled laboratories recognize that information may be incomplete and proactively gather additional data as needed, accounting for internal and external disturbances that might otherwise prevent conditions from being met.
image file: d5dd00151j-f4.tif
Fig. 4 Key differences between conventional laboratory automation (as is) and a SeM-enabled laboratory (to be). Conventional laboratory automation focuses on accurately following user instructions with complete and pre-verified information, typically within the scope of a single or multiple robots. By contrast, a SeM-enabled laboratory treats the entire lab as an integrated system, assumes incomplete information from the outset, and adapts by gathering data as needed. It also takes internal and external disturbances into account, rather than presuming all conditions are satisfied beforehand.

These differences suggest that, in SeM-enabled laboratories, the roles of users and the system are redefined. In existing experimental automation systems, users are responsible for collecting and providing information, and the system determines its actions based on the information provided. Consequently, such systems had to rely on human care for various tasks required to complete an experimental order, such as designing and implementing new experimental protocols, responding to changing conditions, and handling errors. In contrast, SeM-enabled laboratories free users from the need to know the internal conditions of the laboratory when making experimental orders. Instead, the SeM-enabled laboratory autonomously collects the necessary information to fulfill the order—without waiting for user instructions—, recognizes the current state, adapts its responses accordingly, and executes automated experiments.

Similarly, tasks related to higher-level experiment management, such as adjusting schedules or adding new experiments, which were previously the user's responsibility, are now handled by the laboratory itself in SeM-enabled laboratory. For instance, in the case of a schedule change, if one of the sample plates is dropped, the user may be asked whether to cancel the experiment or to continue with the remaining plates. Depending on the user's decision, if the experiment is aborted, the laboratory will clean up and discard the samples. If it is continued, unnecessary steps in the protocol (i.e. the subsequent planned experimental steps for the discarded sample plate) will be omitted to adapt to the situation. Additionally, for new experimental orders, the laboratory autonomously manages and adjusts the schedule. For example, it can use available device time to initiate another cell culture or reallocate operations to accommodate a sudden visitor. Noted that logging critical decision points, criteria, and outcomes, coupled with interfaces allowing human intervention or confirmation, ensures that the system remains visibly under human oversight. This transparency facilitates an incremental transition from human-intensive management to greater autonomy, thus systematically building trust in the SeM-enabled system.

4 SeM-enabled laboratories implementation example

There are several possible approaches to implement SeM-enabled laboratories, but we believe it is important to enable an automated laboratory to “the ability to recognize the conditions in the laboratory and adapt its behavior accordingly.” Existing experiment automation systems require users to perform care to conduct an experiment, because they recognize or assume conditions and decide on actions. By providing the ability to automated laboratories, we expect to eliminate the causes of care and comprehensively remove the burden from the users.

To realize this ability, it is necessary to redesign the devices and its control system for automated laboratories. First, a mechanism is needed to actively gather additional information as necessary, such as details of experimental conditions, sample priorities, or deadlines, because users are assumed to be unable to fully convey their requirements when making an experimental order. Second, a system that actively collects information about the laboratory's state through sensors is essential. Users often lack complete information about the automated laboratory, such as available equipment, labware, or ongoing experiments. Moreover, the laboratory's state can fluctuate constantly due to internal and external disturbances. Third, it is required to flexibly control the automated experiment systems and robots based on the recognized states, demanding the robot controllers to respond to on-demand motion commands rather than executing predefined actions such as teaching and playback. Fourth, while conventional systems treat sensor data and human input separately, a software architecture is required to handle both sources as they jointly influence experimental changes. Fifth, all information in the laboratory—from users to devices to labware—must always be available, even during experiment execution, to ensure the system can adapt to constantly changing states and accommodate new experimental orders or schedule changes.

Hereafter, we will consider the architecture to realize the SeM-enabled laboratory in more detail. We believe that three modules (Requirement manager, Labware manager, and Device manager) and a Central manager that controls them are necessary to realize the SeM-enabled laboratory (Fig. 5).


image file: d5dd00151j-f5.tif
Fig. 5 An example architecture of management software for SeM-enabled laboratory. A laboratory with SeM is referred to as a SeM-enabled laboratory. The system is composed of the Central manager and three key modules: the Requirement manager, the Device manager, and the Labware manager.

The Requirement manager is a module that communicates interactively with the user to obtain and manage the requirements of an experimental order. To conduct a new experiment in conventional experimental automation systems, a user must (1) create a new protocol with detailed specifications, (2) verify that the protocol achieves the expected behavior using empty labware when necessary, and (3) confirm that the required conditions are met at the start of the protocol before initiating it. In a SeM-enabled laboratory, on the other hand, the SeM-enabled laboratory autonomously plans and determines the details of the experiment with reference to the state of the laboratory. Because it is not assumed that users can state all the information required to complete the experimental order at once, the Requirement manager proactively collects additional information on the requirements for the orders through an interactive interface, such as chat.

The Labware manager is a module that manages the current status of the laboratory by recognizing the state of the laboratory through sensing to deal with disturbances that create differences between the assumed and actual laboratory. Sensing is essential in physical space where external disturbances cannot be eliminated, making it impossible to predetermine the state of all objects. Even if a part of the experimental devices is isolated from external factors, some parts of the laboratory must remain open to external interactions to facilitate the entry and exit of consumables. Consequently, external disturbances may occur, such as misplaced consumables or variations in consumable shapes due to lot-to-lot variability. Image recognition is considered useful for recognizing conditions in the laboratory. Recent studies, such as You Only Look Once (YOLO)29 and ot2eye,30 which applied YOLO to labware recognition, can be used to realize SeM. Experimental systems also need the ability to estimate the current state while using past observations. For example, even if a tube is now out of the camera's field of view, it should remain at the last observed position, or the amount of reagent should have decreased since it has been dispensed.

The Device manager is the module that translates the actions determined by the Central manager into specific instructions for the robot or device. The Device manager encompasses device-specific programs to control various devices and robots in a SeM-enabled laboratory. Depending on the state recognized by the Labware manager, the Central manager determines the actions required to achieve a desired state. The Device manager must be flexible enough to adapt to such instructions by the Central manager. While conventional solutions such as dedicated application programming interfaces (APIs) can specify locations in an automated workstation using symbolic position specification. However, in a SeM-enabled laboratory, where multiple devices and robots are used, such solutions as they complicate cross-device coordination (e.g., object transfer and liquid handling across equipments) and make it difficult to handle unexpected situations caused by external disturbances (e.g., misplaced consumables). To address this, the Device manager requires parametric instructions that can flexibly specify positions in physical space, enabling coordination across devices and adjustments for unexpected states. Recently, motion generation AI for robots such as RT-2 (ref. 31) has been studied, and it is expected that this technology can be applied to SeM-enabled laboratories.

These three modules are controlled from a higher level module, the Central manager, which is intended to be a generative AI. The Central manager uses information from users and sensors through the Device manager and the Labware manager to design and proceed experiments through a trial and error process. Trial and error here is not solely mechanical (experiment operation in Appendix B) but also includes resource management, such as allocation of consumables and adjustment of equipment usage time with other experiments (experiment design in Appendix B). As shown in Appendix A, in order to conduct an experiment, it is necessary to determine the experimental design by estimating the necessary resources and repeatedly checking and reserving resources several times. Until now, these tasks have been performed by humans, but with a SeM-enabled laboratory, these tasks are performed by the Central manager.

A SeM-enabled laboratory is realized through the coordinated operation of four modules: the Requirement manager, Labware manager, Device manager, and Central manager. To illustrate the differences between conventional and SeM-enabled laboratories, consider the following scenario. In conventional automated systems equipped with an automated pipetting machine, if a user wants to “add 2 mL of cell culture medium to all wells of this plate,” they must perform the following care: (1) check the available time of the pipetting machine, (2) make a reservation, (3) prepare medium-containing tubes, (4) place the tubes on the deck of the pipetting machine together with the plate, and (5) create and execute an operation program that matches the placement of the tubes.

In contrast, in a SeM-enabled laboratory, when a user requests, “add 2 mL of cell culture medium to all wells of this plate” via the Requirement manager, the Central manager first collaborates with the Labware manager to reserve an automated pipetting machine and medium-containing tubes. Next, the Central manager uses the Device manager to operate a robotic arm to move the plate and tubes to the automated pipetting machine. Finally, the Central manager operates the automated pipetting machine, again via the Device manager, to transfer medium from the tube to the plate. In this way, SeM-enabled laboratories can automate tasks that were previously performed by human operators. More implementation details are provided in Appendix B.

5 Conclusion and discussion

The automation of diverse human tasks in laboratories has long been recognized as essential for stabilizing experimental quality, reducing costs, and improving efficiency. Over the years, various design concepts and implementation strategies have been proposed to advance laboratory automation. These include the development of specialized equipment for automating complex experimental operations and the use of general-purpose robots to handle diverse laboratory tasks.32–34 To automate sample transfer between instruments, mobile robots and rail-mounted robotic arms have been employed.35,36 Proposals for standardized levels and performance metrics of laboratory automation have further clarified goals and benchmarks for achieving greater integration and efficiency.14,37,38 Frameworks for integrating mobile robots with device networks in physical spaces39 and unified APIs for device control40 and inter-device communication41 have been proposed to enhance device coordination. Hierarchical software architectures based on system modeling have also been introduced to control highly automated laboratories.42 Commercial solutions now provide orchestration software that connects multiple devices with mobile robots.43,44 Simplifying and automating experimental protocol descriptions have been enabled by programming languages for abstract protocol design.45 Furthermore, in fields such as cell engineering, the importance of addressing internal and external disturbances in automated laboratories has been emphasized.46 Innovations include digitization and automation of inventory management to ensure seamless resource availability47 and advancements in laboratory information management systems to streamline data flow and workflow coordination.48 Additionally, data representation and exchange schemes have been developed to facilitate efficient communication and interoperability between systems.24,49,50 Automating experiment scheduling51 and enabling iterative optimization by integrating algorithms like Bayesian optimization, which adapt experimental plans based on results, have also contributed to these advancements.14 The vision of fully automated laboratories, capable of solving global challenges, remains a shared aspiration for humanity.2 Despite these advancements, fundamental gaps remain in automating critical human tasks to ensure compatibility, resource availability, and workflow feasibility, including planning tasks (e.g., translating objectives into workflows, scheduling resources) and operational tasks (e.g., experimental setup, monitoring, and error handling).

In this paper, we addressed the critical yet often overlooked concept of care—the human-managed tasks essential for maintaining and supporting automated systems. By proposing SeM as a key enabler of fully automated laboratories, we redefined the automated laboratory as a holistic, adaptive system capable of managing dynamic environments and responding to unexpected disturbances without human intervention. SeM enables laboratories to internalize tasks conventionally performed by humans—such as resource management, error handling, and workflow adaptation—bridging the gap between human intentions and the practical execution of automated experiments. The concept of SeM-enabled laboratories marks a paradigm shift, treating the entire laboratory as a unified, self-regulating entity. We also explored the technical considerations necessary to redesign devices and their control systems to achieve SeM-enabled laboratories. We believe SeM-enabled laboratories not only eliminate bottlenecks stemming from human dependencies but also extend the capabilities of automation to encompass complex, flexible, and diverse experimental workflows, paving the way for a new era of autonomous scientific discovery.

SeM is also crucial for the application of AI-based scientific research, which has recently attracted attention in the field of machine learning, to the biological and scientific fields. Lu et al. recently developed “AI Scientist” and demonstrated the automation of science in the field of machine learning.52 The AI Scientist automates the whole processes of machine learning research including knowledge retrieval, hypothesis generation, validation experiments, data analysis, paper writing, and peer review. The machine learning field is a good benchmark for automation in science as an early demonstration because the experimental process is completed inside the computer. On the other hand, applying this kind of framework to experimental science fields such as life science and chemistry presents a unique challenge. It requires enabling AI systems to autonomously execute experimental processes in the physical space without human intervention or care. SeM-enabled laboratories are fundamental to overcoming this challenge, providing the infrastructure necessary to realize fully autonomous experimental research.

Data availability

This article does not involve any new datasets, code, or analysis.

Author contributions

Conceptualization, K. O.; investigation, all authors; writing – Original Draft, K. O., Y. T., A. K., G. N. K., H. O.; visualization, K. O., K. K., G. N. K.; supervision, K. T., G. N. K, H. O.; project administration, K. O., A. K.; funding acquisition, A. K., K. T., G. N. K., H. O.

Conflicts of interest

K. O., K. K., and K. T. are inventors on patents and patent applications owned by RIKEN covering self-maintainability. H. K. is an employee of YASKAWA Electric Corporation. M. U. is an executive of BAIKEIDO LLC. All the other authors declare no competing interests.

Appendix

A. Flow of executing a specific protocol in the conventional automated laboratory


image file: d5dd00151j-u1.tif

B. Implementation example details

In this implementation example, we propose a departure from the conventional robot control framework to an implementation strategy based on the recently developed Visual Language Action (VLA) model in order to realize the Self-maintainability (SeM) function of the laboratory. Conventional laboratory robots rely on teaching-and-playback and point-to-point control, which lack flexibility and adaptability. In contrast, dynamic control using the data-efficient VLA model has the potential to create an architecture that can respond to complex and variable laboratory environments.
Advancement of VLA models. In this implementation example, the building blocks of the robot control will be redesigned based on the VLA model. Several factors have made this feasible. First, inexpensive hardware such as cameras and robot arms are now readily available. Second, the accuracy and data efficiency of the VLA model have improved significantly. For example, Zhao et al. trained six precise manipulation tasks in the real world from 10-minute or 50 human demonstration trajectories, achieving 80–90% success rate. The learning time was about 5 hours on a single GPU and the inference time was 0.01 seconds.53

For training, teleoperation data is recorded where a person operates the robot in a joint space. During operation, the learned model outputs control signals at 50 Hz, enabling precise movements such as opening a translucent condiment cup and slotting a battery, for example.

Since 2023, the potential for general-purpose robots powered by large-scale models has expanded rapidly. With the advent of VLA models, it is becoming increasingly realistic for robots to autonomously operate while observing their environment, based on natural language instructions—eliminating the need for task-specific programming that was previously required.31,54–58

Application of VLA models in experiment automation. The VLA Model provides an effective foundation for experiment automation. By utilizing developments in these areas as much as possible, cost savings can be expected. On the other hand, experiment automation also requires experiment-specific behavior and the ability to plan at higher layers, such as the design of the experiment.

In traditional programming, the behavior of the system is determined by the architecture, but in machine learning, the dataset design determines the behavior. For example, Chain of Thought is realized by devising training data, not by special software structures59.

This chapter discusses what kind of datasets are necessary for experiment automation from two perspectives: (1) experiment operation: the process of generating behavior based on experimental procedure, and (2) experiment design: the process of creating experimental procedures.


(1) Experiment operation. The experiment operation in this paper represents the process of receiving the experimental procedure and generating the behaviors of the experiment based on it. The behavior here is a sequence of joint angle targets conditioned on the experimental procedure and the sensor values observed at each control cycle time. It is necessary to construct data sets for a variety of tasks, including the following:

● Labware handling: dataset to build on general object manipulation knowledge, but also to learn lab-specific precautions (e.g., keeping plates, tubes, pipettes, etc. horizontal, grasping the top of tube to prevent heat transfer).

● Error handling: collect detection and recovery actions, such as re-grasping when a grasp is missed or interrupting and calling someone if a sample is dropped.

● Replenishment/disposal request: a data set that includes the process from requesting a person to replenishment/disposal, to receiving/removing the item, checking it, and re-requesting it if necessary.

● Use of operator reports: a data set of processes that dynamically adjusts execution plans based on reports such as “reagent refilled”.

● Inventory check: a dataset that includes a series of steps from using a mobile robot's camera to check the consumable shelf and incubator to determine the current inventory.

● Long-term experiment management: the robot is trained to accurately continue a multi-step procedure based on the execution history. Evaluate whether the procedure can be executed without skipping steps or repeating the same part of the procedure.


(2) Experiment design. In this chapter, experiment design is the process of generating an experimental procedure as a preliminary step to an experiment operation to be performed based on user instructions. In the implementation example, when a user gives an experimental instruction such as “Re-execute the cell differentiation procedure from last month with different reagent volume settings,” the system does not immediately start the operation but first constructs a plan step by step by means of a chain of thought. This chain of thought corresponds to the experimental procedure. The experimental procedure is assumed to be something like a Standard Operating Procedure (SOP) written in natural language. In this planning phase, it is necessary to generate flexible procedures according to the situation, referring to the current laboratory environment, past experiment history, and protocols in reference papers.

In particular, in the case of experiments, the following context-adaptive procedure generation is required:

● Description of transportation routes consistent with the physical placement of consumables

● Adjustment of task timing based on available equipment and their free time slots

● Reproduction of past experiments

To improve the accuracy of planning based on these factors, a dataset is needed that enables evaluation not only of adherence to experimental instructions but also from the following perspectives:

● Determining whether the procedure is feasible based on the laboratory's state, including consumables and reservation status for the procedure

● Adherence to standard protocols (e.g., those described in publications or commercial kits)

● Adherence to procedures previously executed in the same laborator

To enable such planning, interactive procedure generation using a VLA model is required. The VLA model generates executable and optimized procedures by referring, as needed, to the current laboratory state, available equipment, past experimental results, and literature, while also confirming the user’s intentions when necessary. For training the VLA model, a dataset that captures the entire flow of procedure creation should be constructed—for example, logs that record both chat history between researchers and technical staff and the corresponding access history to resource management systems and reference materials. This would allow users to avoid designing procedures from scratch; instead, by simply conveying the differences from a published protocol or a past experiment, the VLA model could adjust the procedure accordingly.

In the process of procedure creation, the required quantities of consumables are estimated, and reservations or allocations are made accordingly. If stock is insufficient, actions such as additional ordering are autonomously incorporated into the plan. This process could likely be learned through self-supervised learning, provided that the task of determining whether a procedure is feasible based on the laboratory's state—including consumables and reservation status—has already been achieved.

Furthermore, it is important to establish a mechanism for structuring environmental information in the laboratory so that it can be utilized during experiment planning and execution. Traditional automation systems have had difficulty directly acquiring environmental states, often relying instead on inference from operation history. In contrast, the VLA model can directly perceive the environment through modalities such as vision, allowing the robot to manage state explicitly by writing information back into the environment.

Specifically, the following approaches are taken:

● Attaching notes or labels to labware to manage its contents and status

● Designing rules and placement guidelines for these labels during the procedure generation stage, so that robots can perform the labeling during execution

● Incorporating such “writing to the environment” actions into the dataset from both the experimental operation and experimental design perspectives

Experimental design—generating flexible and explainable protocols while referencing both instructions and the environment—is one of the central elements for realizing SeM.

For implementation, we expect that starting with the automation of experiment operations, which have short individual durations, will facilitate data collection, learning, and evaluation. Furthermore, through this process, it will become possible to identify procedure formats that are easier for the VLA model to translate into operation, allowing the experiment design process to be adjusted accordingly to fit those specifications.

Acknowledgements

This study was supported by grants from the JST-Mirai Program (JPMJMI18G4 and JPMJMI20G7 to K. T. and H. O.), JST BOOST (JPMJBS2414 to Y. T.), Grant-in-Aid for Early-Career Scientists (JP22K17992 to H. O.), Grant-in-Aid for Transformative Research Areas (A) (JP23H04149 to H. O.), Grant-in-Aid for JSPS Fellows (JP22KJ3148 to A. K.), and Grant-in-Aid for Scientific Research (C) (JP23K11820 to G. N. K.). We thank T. Mitsuyama (AIST) and Y. Tanaka (TEGAMI Robotics) for their insightful discussion in the initial meeting; L. Valeggia, M. R. Licheri, J. Meredith, and Y. Kido (TECAN Group Ltd) for invaluable discussions. G. N. K. is supported by Medical Research Center Initiative for High Depth Omics, Nanken-Kyoten, and Multilayered Stress Diseases (JPMXP1323015483), Science Tokyo. We also thank all the lab members at RIKEN BDR, especially R. Nakanishi for insightful discussion. The icons used in this paper were provided by Freepik, dreamicons and iconixar from Flaticon under the free Flaticon License.

References

  1. A. Sparkes, W. Aubrey, E. Byrne, A. Clare, M. N. Khan, M. Liakata, M. Markham, J. Rowland, L. N. Soldatova, K. E. Whelan, M. Young and R. D. King, Autom. Exp., 2010, 2, 1 Search PubMed.
  2. H. Kitano, npj Syst. Biol. Appl., 2021, 7, 29 Search PubMed.
  3. H. Wang, T. Fu, Y. Du, W. Gao, K. Huang, Z. Liu, P. Chandak, S. Liu, P. Van Katwyk, A. Deac, A. Anandkumar, K. Bergen, C. P. Gomes, S. Ho, P. Kohli, J. Lasenby, J. Leskovec, T.-Y. Liu, A. Manrai, D. Marks, B. Ramsundar, L. Song, J. Sun, J. Tang, P. Veličković, M. Welling, L. Zhang, C. W. Coley, Y. Bengio and M. Zitnik, Nature, 2023, 620, 47–60 CrossRef PubMed.
  4. J. F. Hess, T. A. Kohl, M. Kotrová, K. Rönsch, T. Paprotka, V. Mohr, T. Hutzenlaub, M. Brüggemann, R. Zengerle, S. Niemann and N. Paust, Biotechnol. Adv., 2020, 41, 107537 CrossRef PubMed.
  5. E. Matsumoto, N. Koide, H. Hanzawa, M. Kiyama, M. Ohta, J. Kuwabara, S. Takeda and M. Takahashi, PLoS One, 2019, 14, e0212369 CrossRef PubMed.
  6. X. Jiang, S. Feng, R. Tian, G. Han, X. Jiang, M. Ye and H. Zou, Proteomics, 2007, 7, 528–539 CrossRef PubMed.
  7. M. F. Lopez, B. S. Kristal, E. Chernokalskaya, A. Lazarev, A. I. Shestopalov, A. Bogdanova and M. Robinson, Electrophoresis, 2000, 21, 3427–3440 CrossRef PubMed.
  8. C. G. Burns, D. J. Milan, E. J. Grande, W. Rottbauer, C. A. MacRae and M. C. Fishman, Nat. Chem. Biol., 2005, 1, 263–264 CrossRef PubMed.
  9. D. Huang, B. Ou, M. Hampsch-Woodill, J. A. Flanagan and R. L. Prior, J. Agric. Food Chem., 2002, 50, 4437–4444 CrossRef PubMed.
  10. P. Supply, S. Lesjean, E. Savine, K. Kremer, D. van Soolingen and C. Locht, J. Clin. Microbiol., 2001, 39, 3563–3571 CrossRef PubMed.
  11. N. Yoshimoto, A. Kida, X. Jie, M. Kurokawa, M. Iijima, T. Niimi, A. D. Maturana, I. Nikaido, H. R. Ueda, K. Tatematsu, K. Tanizawa, A. Kondo, I. Fujii and S. ’ichi Kuroda, Sci. Rep., 2013, 3, 1191 CrossRef PubMed.
  12. R. Lehmann, J. C. Severitt, T. Roddelkopf, S. Junginger and K. Thurow, J. Lab. Autom., 2016, 21, 439–450 CrossRef PubMed.
  13. E. Vorberg, H. Fleischer, S. Junginger, H. Liu, N. Stoll and K. Thurow, SLAS Technol., 2016, 21, 682–692 CrossRef PubMed.
  14. G. Tom, S. P. Schmid, S. G. Baird, Y. Cao, K. Darvish, H. Hao, S. Lo, S. Pablo-García, E. M. Rajaonson, M. Skreta, N. Yoshikawa, S. Corapi, G. D. Akkoc, F. Strieth-Kalthoff, M. Seifrid and A. Aspuru-Guzik, Chem. Rev., 2024, 124(16) DOI:10.1021/acs.chemrev.4c00055.
  15. J. M. Granda, L. Donina, V. Dragone, D.-L. Long and L. Cronin, Nature, 2018, 559, 377–381 CrossRef PubMed.
  16. M. HamediRad, R. Chao, S. Weisberg, J. Lian, S. Sinha and H. Zhao, Nat. Commun., 2019, 10, 5150 CrossRef PubMed.
  17. B. Burger, P. M. Maffettone, V. V. Gusev, C. M. Aitchison, Y. Bai, X. Wang, X. Li, B. M. Alston, B. Li, R. Clowes, N. Rankin, B. Harris, R. S. Sprick and A. I. Cooper, Nature, 2020, 583, 237–241 CrossRef PubMed.
  18. R. Shimizu, S. Kobayashi, Y. Watanabe, Y. Ando and T. Hitosugi, APL Mater., 2020, 8, 111110 CrossRef.
  19. K. Ochiai, N. Motozawa, M. Terada, T. Horinouchi, T. Masuda, T. Kudo, M. Kamei, A. Tsujikawa, K. Matsukuma, T. Natsume, G. N. Kanda, M. Takahashi and K. Takahashi, SLAS Technol., 2021, 26, 209–217 CrossRef PubMed.
  20. G. N. Kanda, T. Tsuzuki, M. Terada, N. Sakai, N. Motozawa, T. Masuda, M. Nishida, C. T. Watanabe, T. Higashi, S. A. Horiguchi, T. Kudo, M. Kamei, G. A. Sunagawa, K. Matsukuma, T. Sakurada, Y. Ozawa, M. Takahashi, K. Takahashi and T. Natsume, Elife, 2022, 11, e77007 CrossRef PubMed.
  21. R. B. Canty, B. A. Koscher, M. A. McDonald and K. F. Jensen, Digital Discovery, 2023, 2, 1259–1268 RSC.
  22. T. Dai, S. Vijayakrishnan, F. T. Szczypiński, J.-F. Ayme, E. Simaei, T. Fellowes, R. Clowes, L. Kotopanov, C. E. Shields, Z. Zhou, J. W. Ward and A. I. Cooper, Nature, 2024, 635, 890–897 CrossRef PubMed.
  23. K. Darvish, M. Skreta, Y. Zhao, N. Yoshikawa, S. Som, M. Bogdanovic, Y. Cao, H. Hao, H. Xu, A. Aspuru-Guzik, A. Garg and F. Shkurti, Matter, 2025, 8, 101897 CrossRef.
  24. S. H. M. Mehr, M. Craven, A. I. Leonov, G. Keenan and L. Cronin, Science, 2020, 370, 101–108 CrossRef PubMed.
  25. J. Zhou, M. Luo, L. Chen, Q. Zhu, S. Jiang, F. Zhang, W. Shang and J. Jiang, Digital Discovery, 2025, 4, 636–652 RSC.
  26. D. A. Boiko, R. MacKnight, B. Kline and G. Gomes, Nature, 2023, 624, 570–578 CrossRef PubMed.
  27. R. J. Hickman, M. Sim, S. Pablo-García, G. Tom, I. Woolhouse, H. Hao, Z. Bao, P. Bannigan, C. Allen, M. Aldeghi and A. Aspuru-Guzik, Digital Discovery, 2025, 4, 1006–1029 RSC.
  28. N. J. Szymanski, B. Rendy, Y. Fei, R. E. Kumar, T. He, D. Milsted, M. J. McDermott, M. Gallant, E. D. Cubuk, A. Merchant, H. Kim, A. Jain, C. J. Bartel, K. Persson, Y. Zeng and G. Ceder, Nature, 2023, 624, 86–91 Search PubMed.
  29. J. Redmon, S. Divvala, R. Girshick and A. Farhadi, arXiv, 2015, preprint, arXiv:1506.02640,  DOI:10.48550/arXiv.1506.02640.
  30. K. Tachibana and H. Ozaki, ot2eye, Github, 2024 Search PubMed.
  31. A. Brohan, N. Brown, J. Carbajal, Y. Chebotar, X. Chen, K. Choromanski, T. Ding, D. Driess, A. Dubey, C. Finn, P. Florence, C. Fu, M. G. Arenas, K. Gopalakrishnan, K. Han, K. Hausman, A. Herzog, J. Hsu, B. Ichter, A. Irpan, N. Joshi, R. Julian, D. Kalashnikov, Y. Kuang, I. Leal, L. Lee, T.-W. E. Lee, S. Levine, Y. Lu, H. Michalewski, I. Mordatch, K. Pertsch, K. Rao, K. Reymann, M. Ryoo, G. Salazar, P. Sanketi, P. Sermanet, J. Singh, A. Singh, R. Soricut, H. Tran, V. Vanhoucke, Q. Vuong, A. Wahid, S. Welker, P. Wohlhart, J. Wu, F. Xia, T. Xiao, P. Xu, S. Xu, T. Yu and B. Zitkovich, arXiv, 2023, preprint, arXiv:2307.15818,  DOI:10.48550/arXiv.2307.15818.
  32. N. Yachie, Robotic Biology Consortium and T. Natsume, Nat. Biotechnol., 2017, 35, 310–312 CrossRef PubMed.
  33. Y. Jiang, H. Fakhruldeen, G. Pizzuto, L. Longley, A. He, T. Dai, R. Clowes, N. Rankin and A. I. Cooper, Digital Discovery, 2023, 2, 1733–1744 RSC.
  34. K. Darvish, M. Skreta, Y. Zhao, N. Yoshikawa, S. Som, M. Bogdanovic, Y. Cao, H. Hao, H. Xu, A. Aspuru-Guzik, A. Garg and F. Shkurti, arXiv, 2024, preprint, arXiv:2401.06949,  DOI:10.48550/arXiv.2401.06949.
  35. S. Jung, J. Ochs, M. Kulik, N. König and R. H. Schmitt, Proced. CIRP, 2018, 72, 1245–1250 CrossRef.
  36. J. Huang, H. Liu, S. Junginger and K. Thurow, SLAS Technol., 2024, 100240 Search PubMed.
  37. H. G. Martin, T. Radivojevic, J. Zucker, K. Bouchard, J. Sustarich, S. Peisert, D. Arnold, N. Hillson, G. Babnigg, J. M. Marti, C. J. Mungall, G. T. Beckham, L. Waldburger, J. Carothers, S. Sundaram, D. Agarwal, B. A. Simmons, T. Backman, D. Banerjee, D. Tanjore, L. Ramakrishnan and A. Singh, Curr. Opin. Biotechnol., 2023, 79, 102881 CrossRef PubMed.
  38. A. Angelopoulos, J. F. Cahoon and R. Alterovitz, Sci. Robot., 2024, 9, eadm6991 CrossRef PubMed.
  39. Á. Wolf, D. Wolton, J. Trapl, J. Janda, S. Romeder-Finger, T. Gatternig, J.-B. Farcet, P. Galambos and K. Széll, SLAS Technol., 2022, 27, 18–25 CrossRef PubMed.
  40. R. P. Wierenga, S. M. Golas, W. Ho, C. W. Coley and K. M. Esvelt, Device, 2023, 1(4) DOI:10.1016/j.device.2023.100111.
  41. H. Bär, R. Hochstrasser and B. Papenfub, J. Lab. Autom., 2012, 17, 86–95 CrossRef PubMed.
  42. S. Neubert, T. Roddelkopf, B. Göde, X. Gu, S. Junginger, N. Stoll and K. Thurow, in Informatics in Control, Automation and Robotics, Springer International Publishing, 2020, pp. 376–391 Search PubMed.
  43. Biosero, GBG Orchestrator Software Suite, https://biosero.com/software/green-button-go-orchestrator/, accessed 4 January 2025 Search PubMed.
  44. Retisoft Inc., Lab Equipment and Driver Scheduling Automation Software System, https://retisoft.com/genera-automatic-scheduling-software/, accessed 4 January 2025 Search PubMed.
  45. S. Steiner, J. Wolf, S. Glatzel, A. Andreou, J. M. Granda, G. Keenan, T. Hinkley, G. Aragon-Camarasa, P. J. Kitson, D. Angelone and L. Cronin, Science, 2018, 363(6423), eaav2211 CrossRef PubMed.
  46. M. Kino-oka, M. Mizutani and N. Medcalf, Cell Gene Ther. Insights, 2019, 5, 1347–1359 CrossRef.
  47. S. D. Rihm, Y. R. Tan, W. Ang, M. Hofmeister, X. Deng, M. T. Laksana, H. Y. Quek, J. Bai, L. Pascazio, S. C. Siong, J. Akroyd, S. Mosbach and M. Kraft, SLAS Technol., 2024, 1(3) DOI:10.1016/j.slast.2024.100135.
  48. S. G. Higgins, A. A. Nogiwa-Valdez and M. M. Stevens, Nat. Protoc., 2022, 17, 179–189 CrossRef PubMed.
  49. B. A. Schäfer, D. Poetz and G. W. Kramer, J. Lab. Autom., 2004, 9, 375–381 CrossRef.
  50. J. Bai, L. Cao, S. Mosbach, J. Akroyd, A. A. Lapkin and M. Kraft, JACS Au, 2022, 2, 292–309 CrossRef PubMed.
  51. R. Schäfer, J. Assoc. Lab. Autom., 2004, 9, 382–397 CrossRef.
  52. C. Lu, C. Lu, R. T. Lange, J. Foerster, J. Clune and D. Ha, arXiv, 2023, preprint, arXiv:2408.06292,  DOI:10.48550/arXiv.2408.06292.
  53. T. Z. Zhao, V. Kumar, S. Levine and C. Finn, arXiv, 2023, preprint, arXiv:2304.13705,  DOI:10.48550/arXiv.2304.13705.
  54. D. Driess, F. Xia, M. S. M. Sajjadi, C. Lynch, A. Chowdhery, B. Ichter, A. Wahid, J. Tompson, Q. Vuong, T. Yu, W. Huang, Y. Chebotar, P. Sermanet, D. Duckworth, S. Levine, V. Vanhoucke, K. Hausman, M. Toussaint, K. Greff, A. Zeng, I. Mordatch and P. Florence, arXiv, 2023, preprint, arXiv:2303.03378,  DOI:10.48550/arXiv.2303.03378.
  55. M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang and C. Finn, arXiv, 2024, preprint, arXiv:2406.09246,  DOI:10.48550/arXiv.2406.09246.
  56. Gemini Robotics Team, S. Abeyruwan, J. Ainslie, J.-B. Alayrac, M. G. Arenas, T. Armstrong, A. Balakrishna, R. Baruch, M. Bauza, M. Blokzijl, S. Bohez, K. Bousmalis, A. Brohan, T. Buschmann, A. Byravan, S. Cabi, K. Caluwaerts, F. Casarini, O. Chang, J. E. Chen, X. Chen, H.-T. L. Chiang, K. Choromanski, D. D'Ambrosio, S. Dasari, T. Davchev, C. Devin, N. Di Palo, T. Ding, A. Dostmohamed, D. Driess, Y. Du, D. Dwibedi, M. Elabd, C. Fantacci, C. Fong, E. Frey, C. Fu, M. Giustina, K. Gopalakrishnan, L. Graesser, L. Hasenclever, N. Heess, B. Hernaez, A. Herzog, R. A. Hofer, J. Humplik, A. Iscen, M. G. Jacob, D. Jain, R. Julian, D. Kalashnikov, M. E. Karagozler, S. Karp, C. Kew, J. Kirkland, S. Kirmani, Y. Kuang, T. Lampe, A. Laurens, I. Leal, A. X. Lee, T.-W. E. Lee, J. Liang, Y. Lin, S. Maddineni, A. Majumdar, A. H. Michaely, R. Moreno, M. Neunert, F. Nori, C. Parada, E. Parisotto, P. Pastor, A. Pooley, K. Rao, K. Reymann, D. Sadigh, S. Saliceti, P. Sanketi, P. Sermanet, D. Shah, M. Sharma, K. Shea, C. Shu, V. Sindhwani, S. Singh, R. Soricut, J. T. Springenberg, R. Sterneck, R. Surdulescu, J. Tan, J. Tompson, V. Vanhoucke, J. Varley, G. Vesom, G. Vezzani, O. Vinyals, A. Wahid, S. Welker, P. Wohlhart, F. Xia, T. Xiao, A. Xie, J. Xie, P. Xu, S. Xu, Y. Xu, Z. Xu, Y. Yang, R. Yao, S. Yaroshenko, W. Yu, W. Yuan, J. Zhang, T. Zhang, A. Zhou and Y. Zhou, arXiv, 2025, preprint, arXiv:2503.20020,  DOI:10.48550/arXiv.2503.20020.
  57. K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, S. Jakubczak, T. Jones, L. Ke, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, L. X. Shi, J. Tanner, Q. Vuong, A. Walling, H. Wang and U. Zhilinsky, arXiv, 2024, preprint, arXiv:2410.24164,  DOI:10.48550/arXiv.2410.24164.
  58. Z. Fu, T. Z. Zhao and C. Finn, arXiv, 2024, preprint, arXiv:2401.02117,  DOI:10.48550/arXiv.2401.02117.
  59. J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. Le and D. Zhou, arXiv, 2022, preprint, arXiv:2201.11903,  DOI:10.48550/arXiv.2201.11903.

Footnotes

First authors (equal contribution).
Contact authors (equal contribution).

This journal is © The Royal Society of Chemistry 2025
Click here to see how this site uses Cookies. View our privacy policy here.