SYNC
Synchronized on a common clock
Camera frames, joint state, gripper state, and operator intent are aligned to a single timestamp — no manual post-sync.
Capability · Robot data collection
Every teleoperation is a training example.
Yuri records synchronized demonstrations — operator intent, joint state, gripper state, camera frames, and language labels — on every teleop run. The result is a training-ready dataset, not a video you have to post-process.
Why it matters
Physical AI needs demonstration data on real hardware.
The foundation models that power modern robot learning already exist. What they lack is real-world demonstration data — captured on real robot hardware, with real physics, real grippers, and real camera views — to condition on. Yuri is built to produce that data on every teleop run.
That is the difference between a robot and a data-collection platform. A robot that can be teleoperated is useful. A robot that captures demonstrations well — synchronized, timestamped, language-labeled, and export-ready — is what a Physical AI team can build on.
LAB
Language-labeled
Every episode carries a task description, so policies can be conditioned on natural-language commands.
LR
LeRobot-native
Datasets are exported in LeRobot format and ready to sync to the Hugging Face Hub for training.
MCAP
MCAP + Foxglove
MCAP logs and a Foxglove bridge ship with the platform for visual inspection and debugging.
Contains simulated and teleoperated footage. Shipped model may differ from video.
What gets recorded
Operator intent, robot state, cameras, labels.
Four synchronized streams, captured at the same time on a single clock. That is what a training-ready demonstration looks like.
| Stream | What’s captured | Source |
|---|---|---|
| Operator intent | Leader arm pose or Quest 3/3S hand pose — the “what the person wanted to do” signal | OpenLeader / Quest teleop |
| Robot state | Joint angles, torques, gripper aperture, and (on Mobile) base state — “what the robot actually did” | OpenArm 2+ encoders, gripper, base |
| Camera frames | 2× wrist Arducam RGB + 1× head RealSense depth — “what the robot saw” | Wrist + head cameras |
| Language labels | Task description attached to each episode — “what the task was called” | Telepath recording UI |
Operator intent is not needed at inference — it’s needed so the policy can learn from what the operator meant as well as what the robot did. That distinction is what separates a supervised learning dataset from a monitor recording.
Format
LeRobot-compatible, no reshaping required.
Yuri exports datasets in the LeRobot format used across the open-source robot learning community. The output is training-ready: dataset directory, episode structure, timestamps, image tensors, and a language label for every episode. Nothing has to be reshaped before training starts.
Alongside the LeRobot export, Yuri records MCAP logs and exposes them via a Foxglove bridge, so a research engineer can visually inspect a session, compare camera and joint signals, and diagnose a bad episode before it pollutes the dataset.
LR
LeRobot dataset format
The dataset layout used across the LeRobot ecosystem — supported by the training and evaluation tooling in the Yuri stack.
HF
Hugging Face Hub sync
Datasets can be uploaded directly to the Hub, versioned, and pulled into cloud training jobs.
MCAP
MCAP logs
Full-session logs of every topic — useful for inspection, debugging, and reproducing a session offline.
FOX
Foxglove bridge
Live and recorded views of camera, joint, and gripper signals through the Foxglove visualization tooling.
LeRobot and Hugging Face
Built on the same dataset format the field already uses.
Yuri’s demonstration data is not proprietary. It uses the LeRobot dataset format, syncs to the Hugging Face Hub, and is compatible with the open-source policy stack the Physical AI community already trains against.
That means the demonstrations you capture on Yuri are not locked to Sensori software. They can be trained against with the same tooling used across the field — and, later, published back to the community if that is what your lab wants to do.
OPEN
Open format
LeRobot-format exports, not a closed binary. Inspect them with standard tooling or import them into your own pipeline.
HUB
Sync to the Hub
Upload episodes to the Hugging Face Hub for versioning, sharing, and cloud training.
∞
Grows as you record
Each teleop session adds to a dataset that compounds over time — more episodes, more tasks, better policies.
↔
Works with the policy stack
Designed for VLA, Pi-0.x, X-VLA, SmolVLA, GR00T, and custom language-conditioned policies.
What it enables
Collection is step two. Learning is step three.
The demonstrations feed directly into training and evaluation, on-robot or in the cloud. Here is where to go next.
Capture your first demonstration.
Tell us what task you want to record and which configuration fits — Yuri Desktop or Yuri Mobile. We will follow up with pricing, availability, and a suggested collection plan.