The rig is matched to the task — not the other way round.
Time-synchronization and calibration decide whether a multimodal dataset is trainable at all. Poor sync silently corrupts every observation–action pair. So capture method, sync and calibration are engineering decisions we make per task — and validate before delivery.
Wearable multi-sensor
Head-mounted egocentric camera, motion-tracked gloves, wrist IMU, optional fingertip force. Minimal, mobile, and the way most human-demo → robot-action pipelines (UMI, DexCap, EgoMimic) actually recover trainable actions.
Volumetric multi-camera
A 60+ camera ring gives view-invariant, occlusion-robust 3D hand + object pose that a single egocentric view loses under self-occlusion. We use it to extract precise action tracks — never as the deliverable. If a lighter rig captures your task cleanly, we use the lighter rig.
4DGS extraction
4D Gaussian-splatting reconstructs the dynamic scene so we can extract calibrated multi-view RGB-D and 3D pose from arbitrary angles. It is a processing step. You receive RLDS / HDF5 / LeRobot trajectories, not splats.
Sync & calibration, treated as an acceptance gate
We hold a calibration error budget and validate genlock timestamps on every delivery. We report these per engagement — no marketing constants, just the numbers for your data.
The pipeline
Scope
We agree the task family, target embodiment, environments, modalities and delivery format — and which capture rig fits.
Consent & capture
Contributors consent on the record; we capture with genlock-validated sync and per-camera calibration.
Extract & label
We recover the 7-DoF action (and hand pose / force where needed), attach language and calibration, and retarget toward your action space.
Validate & deliver
Every episode passes the acceptance gate — replay, sync, calibration, label integrity — then ships to your ingest spec with a datasheet.