Experiences

Experiences organizes recording evidence and learning work. Training follows Experience → Loops → Runs → Results; see Core concepts for the distinction between a learning context and a recorded attempt.

Recordings and reviews

Import a supported Humaan recording or complete LeRobot v3.0 dataset. Verify artifacts, inspect attempts, and record outcome, assistance, and review state separately. Reviews preserve history without rewriting original recording artifacts.

Learning sets and dataset preparation

Under Loops → Learning sets, select recordings for a stated goal and explain each inclusion. Saved sets pin recording manifests and review revisions. Full exports include selected recordings; manifest-only exports retain references and review snapshots.

Dataset preparation exports eligible data with train/validation assignments and source mappings. Inspect exclusions and the preparation report before using an export. Preparation does not launch training.

Training and evaluation

Under Loops → Training experiments, create an Experience and Loop, save a task and environment, then pin a Builder version and training configuration in an experiment. Choose local or remote compute when launching.

The current CEM trainer optimizes a joint-feedback controller and evaluates it against a baseline on separate fixed starting states. Inspect metrics and replay before registering a candidate in Deployments.

Boundaries

The current trainer does not consume demonstrations or learning sets. Image-based policy training, Dataset Passport compatibility checks, and automatic deployment-session ingestion into a Loop remain planned. Simulation evaluation does not establish physical performance.