Bender

About

Data from where the hands are.

Bender collects first-person RGB-D data for robot learning, and runs the whole process so that what reaches a lab is validated, not raw.

Who we are

Bender is a data-collection company for robot learning. We record people doing ordinary tasks from their own point of view, and we deliver those recordings in a form a lab can load the day they arrive.

We are engineers who kept running into the same gap. The most useful demonstrations for a manipulation policy are a person's hands doing a real job in a real room, and those are the recordings almost nobody has in quantity. Studio setups are staged by nature, and third-person video does not show what the hands see. So we built collection around the phone people already carry.

How a session is collected

Every project starts with a task specification: the activity, the setting, what must stay in view, what to avoid, and how long a session should run. We recruit contributors for that specification, brief them, and take their consent before recording.

Then they do the task the way they normally would, in their own kitchen or workshop, with the phone recording from their point of view. Nothing is staged in a studio. The mess is the point.

What a delivered session contains

One MCAP file. Inside it: color video at up to 30 frames per second, depth maps from the same device, the camera's pose over time, and raw IMU (accelerometer and gyroscope), with every message timestamped on one clock.

The file carries its own channel schemas, so standard MCAP tooling reads it without anything extra from us. Depth, pose and IMU line up with the color frames without post-hoc alignment.

How quality works

Recording is the easy part. Before a session is accepted we review it: that every stream is present and complete, that timestamps agree across streams, that the hands and workspace stayed in view, and that the task was performed as specified.

Sessions that fail are not delivered. Contributors are paid per accepted session, which keeps their incentives and yours pointed at the same thing: a recording you can use.

How to work with us

Tell us the task you want demonstrated, the setting, and roughly how many sessions you need. We reply with a draft task specification for you to correct and a plan for collection. From there, accepted sessions are delivered as they clear review.

If you would rather record sessions than buy them, say so; the same form works, and contributors are paid per accepted session.

Use the contact form; we reply by email.