Collect. Curate. Train.

Hand-grasping demonstrations, from recording to ACT deployment.

01 · Collect

Hand-grasping demonstrations.

Synchronized camera views, measured state and commanded references.

Head

Left wrist

Right wrist

38.0 minDemonstrations

68,364Frames

117Episodes

80Takes

02 · Curate

Review four camera views. Trim takes. Export to LeRobot.

MILO Curation · Mac app

ReviewCameras and robot state share one timeline.

CurateKeep useful intervals; preserve the source.

ExportPrepare aligned LeRobot episodes.

03 · Train

Build the policy graph, choose data and inspect training runs.

ObserveHead and wrist images plus state history.

PredictACT produces 120 reference steps.

EvaluateSplit whole takes into train, validation and test.

04 · Deploy

Observe again, re-plan and execute through the shared whole-body controller.

Camera + state→ACT→Joint references→Shared controller

Navigation

Wheel and visual odometry, lidar mapping and Nav2 goals enter the shared command path.