BALINA / PRIVATE PREVIEW

See what the model suggests.

SmolVLA observes the real robot’s cameras and joints. Inspect its predictions before adding any model control.

Read-only · no motor commands

One snapshot. One prediction.

Load the robot view, describe a task, then ask SmolVLA. Each request captures fresh observations. The robot must be idle with no active operator.

Operator access key · stays in this tab Task for the model

Enter your operator key to use Balina. Nothing is sent automatically.

What is shared with the model?

The overhead and wrist camera snapshots, six normalized joint readings and your task text are processed on Balina. The third camera input is a black placeholder because this robot has two cameras. Images are resized with padding to 256 × 256. These requests do not save a dataset or start a robot-control session.

The hosted preview is restricted to operators using the existing private robot-console key. The key is checked by VeraPulse and never forwarded to the model. A regular website account does not grant model access.

What the model saw

Captured snapshot · not a live stream

Captured overhead robot camera input — 01 / Overhead

Captured wrist robot camera input — 02 / Wrist

Black placeholder for the missing third camera — 03 / Missing camera · black padding

Joint readings and model output

Joint	Measured ticks	Model input	Model output

Model inputs use the existing calibration: arm joints −100 to 100, gripper 0 to 100. An asterisk means a reading was clipped at the calibration boundary. Outputs are shown as returned by the checkpoint; they are never converted into motor commands.

Experimental base model. Camera mapping and task performance are not validated. A prediction does not establish that a movement is correct or safe to execute.

Model connection guide ↗