Training

Turn uploaded recordings into a trainable dataset, train a policy in the cloud, and run it back locally. Capturing the recordings themselves is covered in Video and recording.

From recordings to a trainable dataset

A recording isn't trainable by itself. You must assemble it from My Stuff, which flattens one or more recordings into a dataset (a new one, or appended onto an existing one). Datasets are LeRobot-format, so existing training pipelines apply without conversion, and each episode records the grippers' current draw (raw and normalized) alongside joint positions — a grip-force signal with no dedicated force sensor. Assembly takes minutes; the dataset is trainable — and downloadable — as soon as it's done.

Synthetic maps and relighting (optional)

At assemble time you can enrich a dataset with derived per-frame maps ( depth, normals, albedo, roughness) generated from the recorded video (depth from a segmentation pass; the other three from a single inverse-rendering pass), for all cameras or just the ones you pick. Two video-processing passes are also offered: color_jitter and full_relight (synthetic relighting).

Two things to know before ticking the boxes:

• Map derivation runs on cloud GPUs after assembly and can take hours for a large selection — cost scales per episode × camera, and the assemble dialog quotes the two phases separately. The dataset itself is trainable immediately; maps arrive later.
• Progress and failures show on the dataset card in My Stuff. A requested map renders muted until it's produced; finished maps are downloadable and selectable in the episode viewer.

Training and inference

Inference is local — which is why the desktop download is ~770 MB rather than ~200 MB. A robot's motor-command loop must not depend on your Wi-Fi holding up.

Training is cloud — it needs GPUs you don't have, and it's not latency-sensitive. Nothing about training touches the robot.