ZEN MATRIX

The operating layer for real-world robot data.

Turn asynchronous MCAP streams into aligned, traceable and model-ready datasets. Zen Matrix connects ingestion, synchronization, annotation, quality control, dataset lineage and export in one governed workflow.

CAPTURE · SYNCHRONIZE · GOVERN · EXPORT

MCAP / RGB / Depth / Joint State / End-Effector Pose / Force-Torque / IMU / Teleoperation

REF · LeRobot · RLDS · HDF5 · Zarr

SEE THE DATA BEFORE YOU TRAIN THE MODEL

Zen Matrix makes every stage of a robot dataset visible: what arrived, what aligned, what was changed, what passed quality control and what was delivered.

PRIVATE PREVIEW · MCAP IN · MODEL-READY DATA OUT


01 · MCAP INGESTION & TOPIC EXPLORER

Upload an MCAP dataset and inspect every Topic, message type, timestamp source, frame count and effective sampling frequency.

  • RGB and depth cameras
  • Joint position, velocity and torque
  • End-effector pose and gripper state
  • Force / torque, IMU and teleoperation commands
  • Calibration, device, firmware and operator metadata

02 · RAW ASYNCHRONOUS TIMELINE

Visualize the original sampling points before processing. Frequency drift, clock offset, gaps, duplicated timestamps and out-of-order messages remain visible instead of being hidden.

RGB CAMERA •----•----•------•----• 30 Hz
JOINT STATE •-•-•-•-•-•-•-•-•-•-• 100 Hz
TELEOP COMMAND •--•--•--•--•--•--• 60 Hz
FORCE / TORQUE •••••••••••••••••••• 500 Hz
TARGET TIMELINE |---|---|---|---|---| 50 Hz

03 · ALIGNMENT WORKBENCH

Configure one target timeline while choosing the correct method for each data type.

  • Continuous joint data: linear or higher-order interpolation
  • Discrete gripper and mode states: forward hold or nearest neighbour
  • Position and orientation: translation interpolation plus quaternion SLERP
  • Images: nearest valid frame, never blind pixel interpolation
  • Maximum time tolerance, gap policy and missing-data flags for every Topic

04 · BEFORE / AFTER COMPARISON

At any target timestamp, compare the camera frame, robot state, end-effector pose, force signal and teleoperation command side by side.

Visible metrics: maximum time error · median offset · missing rate · effective frequency · frame drop · interpolation ratio · confidence score

05 · ANNOTATION & EPISODE ONTOLOGY

Move from synchronized signals to task understanding. Create task, phase, event, object, contact, success / failure, anomaly and skill labels on the same timeline. Automated proposals remain traceable to the model, prompt and reviewer.

Explore the Annotation Workbench →

06 · DATA QUALITY REPORT

  • Timestamp precision and cross-sensor offset
  • Sampling stability, gaps and duplicate messages
  • Trajectory smoothness and action-observation consistency
  • Camera continuity, calibration health and view coverage
  • Episode completeness, anomaly detection and acceptance score

07 · LINEAGE, VERSIONING & DELIVERY

Every dataset version records its source files, transformation parameters, calibration version, annotation revision, quality result and approval history.

EXPORT: REF · LeRobot · RLDS · HDF5 · Zarr · project-specific schema


A GOVERNED DATA PIPELINE FOR EVERY DEPLOYMENT

Zen Matrix can run as a project Demo, a customer-controlled private deployment or the operating layer of a continuously growing robot data factory.