08Data & learning

The path from teleop to autonomy: record episodes while you operate, curate them, train with LeRobot, deploy back onto the robot. Every stage has a browser studio; all of them run on the laptop.

Recording

Episodes — any client (or the studio) sends the record wire message; the gateway captures synchronized camera frames + joint states per episode, organized by task name.
Raw logs — data_collect records rosbag/MCAP on the robot for anything the episode format doesn't carry.
Recording state is visible in every client and in the cockpit's status row (● REC).

The studios

Port	Studio	Purpose
8093	Data Curation	browse/score episodes, drift checks, filter bad takes, export LeRobot datasets
8092	Data Studio	HDF5 episode editor — trim, splice, inspect frame by frame
8095	Trainer	LeRobot training GUI — ACT, diffusion (DiT), pi0; run/queue/compare jobs
8096	Community queue	public training queue with owner approval (optional, VPS-hosted)

Typical loop

Collect 30–100 episodes of one task via Vision Pro or iOS teleop, consistent staging, task-named.
Curate in the Data Curation studio — drop failed takes, check distribution drift, export a LeRobot dataset.
Train in the Trainer studio (ACT is the fastest baseline; DiT and pi0 configs ship alongside). GPU strongly recommended.
Deploy — select the policy from a client; the runner streams actions through the same wire contract and safety gates as teleop. The e-stop and tip-over model apply to policies exactly as to humans.

Simulation & RL

The same MJCF drives RL environments (a vendored mjlab setup with lift/bimanual/whole-body tasks under learning/) and the Genesis/MuJoCo/Unity twins — policies can be shaped in sim against the same wire contract before they ever touch hardware.

Every deployment path — scripted, learned, or teleop — goes through the same single gate (safe_base_step, the tip-over model, self-collision QP). There is no "policy mode" that bypasses safety.