Zen is the collective name for both the intelligence that enables robots to understand and execute real-world manipulation tasks, and the underlying model and data infrastructure that continuously improves that intelligence.
In recent robotics research, new architectures such as Vision-Language-Action (VLA) Models and World Action Models (WAMs) are emerging one after another. However, model architectures and hyperparameters evolve on the scale of months, with new research rapidly becoming available through academic papers and open-source implementations. The models themselves are therefore unlikely to remain a long-term competitive advantage.
What matters is the ability to deploy robots into the real world, transform operational data into trainable assets, evaluate improvements on physical robots, and continuously iterate. At the core of Zen is this improvement loop that remains directly connected to the real world.
The Field Is in Japan
Approaches such as VLA and WAM are attracting tremendous attention, but there is still a considerable gap between research and real-world deployment. Most of today’s work remains confined to the laboratory.
Meanwhile, Japan faces a severe labor shortage, with many real-world problems that people want solved today. We are developing Zen to bridge the gap between cutting-edge robotics research and systems that are genuinely useful in practice.
Validate in the Real World from Day One
Zen’s first implementation did not begin with a laboratory benchmark—it began in a hospital.
In March 2026, we conducted a proof-of-concept trial at the University of Tsukuba Hospital, evaluating nighttime patrols and multilingual guidance services. During this trial, we primarily validated the robot’s communication and navigation capabilities, confirming the use case of a bipedal robot autonomously guiding people to their destinations while avoiding obstacles within a real hospital environment.
In June 2026, we expanded the trial at the same hospital to include manipulation use cases, where the robot performed blood collection preparation tasks.
The system used here was not a single end-to-end model.
A state machine for managing long task sequences
An imitation learning model for precise manipulation skills
A VLA-based post-training model
These components were integrated into a single system to execute real-world tasks.
The operation tasks that make up the blood collection preparation workflow achieved a 7/7 success rate under teleoperation and 3/6 using the trained model, when evaluated in the same facility during the same time period.
External research has also quantitatively demonstrated that optimizing each individual action does not necessarily improve the overall task. In one study, a frozen VLA model was combined with residual reinforcement learning to optimize individual subtasks. Although the success rate of the insertion task alone increased dramatically from 45.7% to 92.2%, the success rate for the complete workflow reached only 54.5%, failing to outperform the end-to-end baseline that performed no task decomposition (54.9%). This occurs because optimizing each subtask independently tends to converge on intermediate states—such as grasp configurations—that are disadvantageous for subsequent steps. A high subtask success rate is therefore not a sufficient condition for successfully completing the overall task.
The difference between 7/7 under teleoperation and 3/6 with the trained model is precisely the gap that Zen is designed to close.
Turning Real-World Experience into Trainable Assets
Zen is not intended to produce robots that succeed only once in a demonstration or video. Its goal is to repeatedly execute tasks reliably in real-world environments that include environmental variation, object variability, unexpected interruptions, and changing contact conditions.
To achieve this, Zen operates the entire process—from data collection to redeployment—as a single continuous pipeline. Video, joint states, and action commands recorded by robots and data collection devices in the field are first synchronized in time and converted into trainable datasets. Every dataset and model is tracked with versioning and lineage information, recording exactly which datasets produced which models. Trained models return to the field only after passing both automated evaluation and validation on physical robots. Every new experience acquired in deployment—whether successful or unsuccessful—becomes input for the next round of training. The faster this cycle runs, the faster the robot becomes more capable.
Converting raw data into datasets also requires extensive validation to ensure data quality. Both automated and manual checks are performed. Examples include verifying that there are no missing frames in video sequences, ensuring that gripper signals have not frozen, checking that wrist-mounted cameras have not been pressed against objects and obscured, and finally conducting human audits.
Before a trained model is returned to the field, the automated evaluation gate consists of open-loop Action MSE measured on a held-out evaluation set. Approximately 10% of the recorded episodes that were not used during training are replayed, and the actions produced by the trained model are compared against the actions originally recorded. Both imitation learning models and post-trained models pass through this same evaluation gate.
For closed-loop evaluation, checkpoint-level validation is performed using simulation environments constructed through real2sim/sim2real techniques. By executing policies within simulated environments that faithfully reproduce the deployment site, we can determine which checkpoint should be promoted before bringing the model onto a physical robot.
The final gate, however, is evaluation on the physical robot. Automated evaluation can quickly and inexpensively verify that performance has not regressed, but ultimately the model must still be validated on real hardware. Because we develop our own hardware platform, D1, we are able to control this entire cycle and iterate much more rapidly.
A Single Data Source Cannot Capture Reality
Zen places real robot data at its core while combining it with human-generated data.
Robot Data
D1 is operated either through teleoperation or autonomously using trained models to collect paired video and action command data directly from physical robots.
This data captures the complexities that exist in production environments, including real-world friction, lighting conditions, object variation, and human movement. Because it is collected in the actual deployment environment, it follows the same distribution as production (in-distribution) and naturally contains long-tail failures that are absent from scripted demonstrations, such as shelves catching unexpectedly or objects shifting during manipulation. Faces, name badges, and documents are masked on-device, and only anonymized data is retained.
Glove-Based Data
In addition to teleoperating D1, we also collect contact-rich manipulation data using dedicated data collection devices such as YUBI and specialized sensor gloves.
The glove is equipped with seven IMUs that measure finger joint angles, along with 162 pressure sensors distributed from the fingertips to the palm. Hand motion and contact information are recorded at 1 kHz. This allows us to preserve information that cannot be captured through video alone—where a person touched an object and how much force they applied—as trainable data.
Scaling data collection using robot teleoperation alone is extremely difficult, because the amount of data that can be collected is constrained by the number of robots and their available operating time. Sensor gloves, on the other hand, are not tied to robot hardware. Data can be collected simply by having people perform tasks using their natural hand movements, and because the devices are significantly less expensive than robots themselves, increasing the number of operators scales data collection in parallel. Naturally, the quality of glove-collected data is lower than that of robot teleoperation data, but this division of roles makes data collection both economical and efficient.
From a Single Field Demonstration to a Proprietary Data Collection Cycle
In addition to collecting robot data through teleoperation and glove-based demonstrations, Zen amplifies robot data using its proprietary Field Shot Learning (FSL) method.
When examining repeated demonstrations of the same task during data collection, most of what actually changes is not the manipulation itself, but the context in which it is performed. What varies are factors such as the robot’s position, the table position, object placement, and lighting conditions. The manipulation itself changes very little. FSL is built around explicitly separating these two aspects.
The robot observes the manipulation only once. Thereafter, each time it executes the task, it measures the environment itself and re-solves the demonstrated behavior to match the current situation. The demonstrated manipulation remains unchanged, while only the parts that should vary according to the environment are adapted.
By recording the task only once in the field, the robot can subsequently continue collecting robot data autonomously while gradually changing the environment in which the task is executed, without requiring further teleoperation. Each re-solved execution becomes a new real-world episode. It captures real friction, real lighting, and real contact conditions, with both successes and failures retained as structurally labeled data. No human operator is required to stand beside the robot. As a result, the amount of data that can be collected is no longer limited by human time, but instead scales with the robot’s operating time.
Competitive Advantage Comes from the Speed of Improvement
New models are being released every day, and new architectures continue to emerge.
The best approach available today may no longer be the best just a few months from now. That is why Zen is not built around any single model. Instead, what we continue to build is the capability to replace models while continuously producing improvements that have been validated in the real world.
What we want to maintain is the ability to deploy the latest robot foundation models into real-world environments faster than anyone else, and to determine as quickly as possible whether they have actually improved performance. Achieving this requires robots that remain continuously deployed in the field, a constant stream of data flowing back from those deployments, and an evaluation framework that rigorously questions the results. Zen is the infrastructure that makes this possible.
Labor shortages will not wait. Rather than waiting for laboratory benchmarks to make their way into the real world, we build solutions that work by starting from the field itself.