☁
Cloud training
Training jobs are provisioned in the cloud from the browser UI. No local GPU required to get started.
Capability · Robot learning
From demonstration to deployment.
Once demonstrations are captured, Yuri moves through the rest of the loop: train a policy in the cloud, evaluate rollouts on the robot, improve the dataset, and deploy the learned skill. The full workflow lives behind one browser screen and is built on the open LeRobot stack.
Workflow
Four steps, one screen, no command line.
Yuri’s learning workflow is built around the loop Physical AI teams actually need: collect demonstrations, train policies, evaluate rollouts, improve the dataset, repeat. Telepath runs the whole thing from any browser — laptop, tablet, phone, even inside a VR headset.
Calibrate
Guided setup finds the arms and cameras and remembers them.
Record
Teleoperate the robot and record synchronized episodes.
Train
Pick a model, pick a GPU, one click. Training runs in the cloud.
Run
A safety checklist, then the robot performs the task on its own.
Training
Pick a model, pick a GPU, one click.
Training runs in the cloud, triggered from the same browser screen that recorded the demonstrations. Choose a dataset, choose a policy, choose a GPU, and start — training is orchestrated off the robot so the hardware stays available for data collection and evaluation.
For teams that prefer to train on their own infrastructure, Yuri’s LeRobot-format datasets import into standard tooling. For teams that prefer to stay inside the Yuri stack, training is provisioned per-minute from the UI, with live pricing shown before the job starts. Both paths produce policies that run on the robot.
$
Per-minute pricing
The start form shows the GPU and its per-minute price before the job runs — no surprise bills.
GR
LeRobot and GR00T N1.7
Supported policies include NVIDIA GR00T N1.7, ACT, pi0.5, and any policy in the LeRobot stack.
W&B
Weights & Biases tracking
Training runs log to Weights & Biases for experiment tracking and comparison.
Evaluation
Rollouts on real hardware, watched by a human.
Evaluating a policy is not the same as training it. A trained policy can look good on validation loss and fail in contact. Yuri’s evaluation loop runs the policy on the robot, in the same workspace the demonstrations were recorded in, with a safety checklist and an emergency stop always available.
Two things matter in evaluation: the rollouts need to be recorded, so the failure modes can be inspected, and they need to be reviewable, so the next round of collection can be targeted at the failure. Yuri does both by default — every rollout is logged in the same MCAP format as the demonstration episodes, and the camera feeds stay visible throughout the run.
R
Rollout recording
Every evaluation run is recorded in the same MCAP and LeRobot formats as the demonstrations.
SH
Shadow mode
New skills run in “shadow mode” first — the robot thinks, but does not move — until motion is deliberately enabled.
E
E-stop always available
A big Emergency Stop is one tap away in the UI and always cuts motor power. A hardware e-stop ships on the robot.
◎
Depth-camera safety
The head-mounted RealSense depth camera can detect people too close to the robot during a rollout.
Deployment
A trained policy runs on the robot, at the edge.
Yuri ships with an NVIDIA Jetson AGX Orin 64 GB preconfigured for on-robot inference — 275 TOPS on the robot, so inference does not depend on a network round trip. Trained policies are deployed to the robot, run in shadow mode first, and then enabled for motion.
The same AGX Orin that runs inference is the compute node that runs teleoperation and recording. That means a team can go from demonstration, to policy, to live autonomy on the same device — with the same software image, the same calibration, and the same safety stack.
275
275 TOPS on-robot
NVIDIA Jetson AGX Orin 64 GB, preconfigured for VLA training and inference on day one.
↺
Signed OTA updates
Software updates ship signed, with rollback — deploy a new policy or a stack update without reflashing the robot.
↔
Desktop or Mobile
The same policy deployment workflow applies to Yuri Desktop and Yuri Mobile, on the same software image.
∞
Iterate
Every rollout adds to the dataset. Every dataset improves the next policy. The loop compounds.
Supported policies
Open-stack policies, or your own.
Yuri is designed for vision-language-action (VLA) policy workflows and for the policy families the Physical AI community already trains against. The shipping stack supports NVIDIA GR00T N1.7, ACT, pi0.5, and any policy in the LeRobot ecosystem — and Yuri is designed for Pi-0.x, X-VLA, SmolVLA, and custom language-conditioned manipulation policies.
GR
NVIDIA GR00T N1.7
Supported in the shipped policy stack, trained and evaluated through the Yuri workflow.
PI
Pi-0.x / pi0.5
Designed for Pi-0.x-style policies; pi0.5 is supported in the shipped stack.
ACT
ACT / SmolVLA / X-VLA
LeRobot-native policies work on Yuri by default; X-VLA and SmolVLA are designed for.
CUSTOM
Custom policies
LeRobot-format datasets and standard ROS 2 / Python integration mean your own policy can run on the robot.
LLM
Language-conditioned
Yuri accepts natural-language task prompts, so trained policies can be selected and conditioned by language.
The full capability set
Teleoperation, data collection, and learning are one loop.
Each capability feeds the next. Teleoperation produces demonstrations. Data collection turns them into training datasets. Learning turns datasets into policies. Then teleoperation starts again, on the failure the policy exposed. Here is where to go next.
Train your first policy on Yuri.
Tell us what task you want to teach and which configuration fits — Yuri Desktop or Yuri Mobile. We will follow up with pricing, availability, and a suggested learning plan.