Train models on your robots’ own data.
Record episodes
Every camera stream and every joint reading, captured on one clock at the rate you declare, synced and uploaded without a logging rig of your own. Teleoperate the arm, mark the episode, and the dataset builds itself.
Config-driven capture from any camera, sensor, or arm, synced to the cloud continuously or on demand
Declared capture: which resources, which methods, what rate, one clock
Teleop with a leader arm or a controller; episodes saved as sequences
Export LeRobot-compatible datasets
Train a model in Viam’s cloud
Every machine captures the images and readings you declare, syncs them to one dataset, and the app labels, trains, and evaluates a detector or classifier on them. The trained model deploys back to the machines as a config line. When the fleet sees something new, the next version trains on that too.
Config-driven capture from any camera or sensor, synced per fleet
Label and train detectors and classifiers in the app; bring your own training script
Deploy the model to one machine or the fleet through the ML model service; versioned, revertible
Retrain on production data as it accumulates
Deploy and run on your robot
Fine-tune a vision-language-action model on your sequences, or pull a checkpoint straight from Hugging Face, and the policy comes back to the machine as one more declaration: a service in config, driving the arm through the same API, under the same limits as everything else. Async chunked inference keeps the arm smooth while the model thinks.
Run a learned policy as a Viam service, with any checkpoint: yours, the Registry's, or Hugging Face's
Inference on the robot's own compute (a Jetson is enough)
Delta and joint-limit clamps enforced below the policy
