Two gloves, three cameras, one clock

2× MANUS gloves <15 ms @ p95 3× RealSense 45 episodes

An actual dataset recording

A teammate at Nferent AI runs the task while two motion-capture gloves read their hand pose and three depth cameras watch — all recorded at once. The hard part is time: getting all five streams to agree on when something happened, so the frame where the fingers close is the frame stored next to it. The sync layer is mine; the collection is the team's.

The gloves, tracked live

The gloves on their own — finger and hand pose mirrored to the on-screen skeleton in real time. One of the five streams my pipeline pulls onto a single clock.

How it works

Two MANUS gloves stream finger and hand pose; three RealSense cameras stream RGB-D from three viewpoints. My software ingests all five streams at once.
It stamps every stream against one shared clock and aligns them on the way to disk. Measured drift between streams stays under 15 ms at the 95th percentile.
Forty-five episodes have been recorded through this pipeline. Loose sync quietly poisons a dataset — a policy can't learn from frames that disagree about when something happened — so the millisecond budget is the whole job.

My part, honestly

To be clear about credit: the data collection is run by a teammate at Nferent AI. What's mine is the capture-and-sync software behind it — ingesting the camera feeds and glove data and holding them to one clock. Nferent AI shared the rig running on their LinkedIn, which is what lets me show it here at all. The failure log for that pipeline is a write-up in progress.

Links

The rig in action, posted by Nferent AI