Research

We build for the robotics community as much as for ourselves.

What we learn gets released, so any team can reproduce it, challenge it, and take it further.

We would rather see a hundred groups make progress here than hold a narrow advantage by ourselves.

When a research team collects better data because of something we built, or a company deploys a robot that would not have worked otherwise, our research has done its job.

That is the mission. Robots that work in the real world, and a community strong enough to keep building them.

Papers

• In-the-Wild Annotation: Spoken Reasoning Traces for Robot Manipulation Data

  In preparation

  The operator narrates while demonstrating and the speech is transcribed and time-aligned into subtask boundaries and step-level reasoning traces at the collection site; because narration is unconstrained the labels are open-vocabulary and cover off-nominal events such as a dropped object and its recovery, and a within-operator narrated-versus-silent comparison measures what narration costs in demonstration quality and collection throughput.

• HandUMI: Modular Robot-Free Capture for Bimanual Arms

  In preparation

  HandUMI is an open-source, low-cost, modular UMI-style hand-worn capture device for fixed-base bimanual arms with parallel-jaw grippers. Detachable fingertips match the target gripper, VR controllers track both wrists, wrist cameras record task observations, and a servo encoder measures jaw aperture directly. The synchronized raw capture stays robot-agnostic until calibration, validation, and retargeting.
  Read the white paper

Back