The most critical capability of robot intelligence is not a static model, but a system that continuously learns, tests, evaluates, and improves directly in the real world. "Zen" is our manipulation intelligence designed to understand, execute, and continuously advance robot manipulation tasks in the field.
In modern robotics research, new architectures like Vision-Language-Action Models (VLA) and World Action Models (WAM) emerge constantly. However, model architectures and hyperparameters are updated every few months and rapidly commoditized through papers and open-source releases. Models alone do not offer long-term competitive advantage.
What truly matters is deploying robots into the real world, converting real-time operational data into trainable datasets, and continuously iterating through physical evaluation. At the core of Zen lies this real-world-connected flywheel.
Japan as the Ultimate Testing Ground
While VLA and WAM architectures are widely hyped, a massive gap remains between lab research and real-world execution. In Japan, severe labor shortages create an urgent demand for immediate, practical solutions. We developed Zen specifically to bridge this gap between cutting-edge technology and real-world field utility.
Validated in the Field from Day One
Zen’s initial deployment started in a real hospital environment rather than standard lab benchmarks.
March 2026 (Tsukuba University Hospital): Verified night-time patrolling, multi-language navigation, and dynamic obstacle avoidance using a bipedal robot.
June 2026 (Tsukuba University Hospital): Expanded into physical manipulation tasks, successfully executing blood draw preparation workflows.
Crucially, this deployment did not rely on a single end-to-end model. Instead, we integrated a hybrid architecture:
State Machines for long-horizon task orchestration
Imitation Learning Models for precision manipulation skills
VLA Models for post-processing and adaptation
Field Performance vs. Task-Level Optimization
For the blood draw preparation task under identical real-world conditions, we achieved:
Teleoperation Success Rate: 7/7
Trained Model Success Rate: 3/6
External research quantitatively demonstrates that optimizing individual sub-tasks fails to improve end-to-end performance. For instance, in Freezing VLA with Residual RL for Sub-task Optimization, sub-task success jumped from 45.7% to 92.2%, yet full-pipeline success capped at 54.5%—failing to beat an un-decomposed end-to-end baseline (54.9%). Optimizing sub-tasks independently creates local optima that harm subsequent steps.
Sub-task success is not a sufficient condition for task completion.
The gap between teleoperation (7/7) and model autonomy (3/6) represents the exact distance Zen is built to close.
Transforming Field Experience into Trainable Assets
Zen does not target one-off demo success. It builds repeatable task execution across variable environments, dynamic objects, unexpected halts, and physical contact shifts.
To achieve this, Zen operates as a single unified pipeline from data collection to redeployment. Multi-camera video, joint commands, and action signals captured by field robots are time-synchronized and converted into trainable datasets. Every dataset and model is managed with strict versioning and lineage tracking. Trained models pass automated evaluations and physical hardware tests before returning to the field. Every field experience—success or failure—becomes the next training input. The faster this iteration loop runs, the faster the robot scales its intelligence.
To guarantee dataset quality, we enforce both automated checks and human auditing: mechanical checks for missing video frames or frozen gripper signals, and verification against severe camera occlusion or extreme contact forces.
Regarding evaluation, the automated gate before returning a model to the field relies on open-loop Action-MSE on held-out episodes (~10% hold-out data). For closed-loop evaluation, we utilize real2sim / sim2real simulation environments for checkpoint validation.
Finally, the ultimate gate is the physical robot. While automated evaluation proves "it hasn't regressed" cheaply and quickly, final validation requires physical hardware. Developing our own D1 hardware enables us to control and rapidly execute this end-to-end cycle.
A Single Data Source Cannot Cover Reality
Zen anchors itself on real physical data while integrating human-derived data.
Robot On-Device Data
Collected via D1 teleoperation or autonomous execution, capturing real-world complexity: friction, lighting shifts, object variation, and long-tail corner cases. Faces, name tags, and documents are automatically masked on-device to strictly process anonymized data.
Tactile Glove Data
To scale data collection beyond physical robot bounds (limited by fleet size and uptime), we collect contact-rich manipulation data using specialized wearable devices like YUBI and custom gloves (equipped with 7 IMUs and 162 pressure sensors at 1kHz). Highly affordable, human-scalable, and decoupled from robot hardware constraints.
Autonomous Data Collection via Field Shot Learning (FSL)
In addition to teleoperation and glove data, Zen employs a proprietary method called Field Shot Learning (FSL) to amplify robot datasets.
During task trials, motion trajectories remain largely invariant; what changes are environmental factors—robot posture, table position, object orientation, and lighting. FSL decouples motion intent from environmental variance.
The robot observes a motion once, perceives environmental shifts prior to each execution, and re-solves the trajectory on the spot. It continuously acquires autonomous robot data under shifting environments without requiring teleoperation. Every re-solved execution yields a structured, labeled episode including real friction, contact, success, or failure.
Competitive Advantage Lies in Iteration Velocity
New models and architectures will continuously emerge, and today's state-of-the-art technique will inevitably be superseded. Zen does not bet on a single model. Our core advantage lies in our systemic ability to continuously generate real-world-validated improvements while swapping out underlying models.
We aim to test new foundational robotics models in the field faster than anyone else and accurately evaluate whether they genuinely improve real-world performance. This requires:
A continuous physical robot presence in the field
An uninterrupted stream of real operational data
Rigorous, ground-truth evaluation gates
Labor shortages will not wait for laboratory research to mature. Instead of waiting for benchmarks to trickle down from academia, we build operational intelligence directly from the ground up in the field.