Skip to content
Bud's Wiki

HandUMI — Robot-free bimanual data collection retargetable to any parallel-gripper arm (Brik H. Meza launch tweet)

TL;DR

Section titled “TL;DR”

Brik H. Meza (@autobrik) announces HandUMI, an Apache-2.0 open-source data-collection interface for bimanual parallel-gripper arms. The pitch: record once with a robot-free wearable capture rig and replay the same demonstrations directly on Agilex PiPER, OpenArm, TRLC-DK1, and I2RT YAM by swapping the gripper. HandUMI outputs LeRobot v3-compatible datasets, ships built-in calibration + QA before conversion, and supports both sim replay and real teleoperation. Framed as an accelerator for both startups (deployment) and researchers (experiments) around the thesis that bimanual arms are the right embodiment to start adding real-world value.

Key claims

Section titled “Key claims”
  • HandUMI is designed for robot-free capture — no robot in the loop during data collection [tweet 1/].
  • Recorded demonstrations replay onto four listed bimanual parallel-gripper arms — Agilex PiPER, OpenArm, TRLC-DK1, I2RT YAM — via a swappable gripper attached to the arm of choice [tweet 1/].
  • Datasets are emitted in LeRobot v3-compatible format, with built-in calibration + QA gates before conversion [tweet 1/].
  • The system supports both sim replay and real teleoperation from the same captured data [tweet 1/].
  • Licensed Apache-2.0 and fully open source [tweet 1/].
  • Co-authors credited: @alvax64, @leoperzz, @raulb4s, @mbrq_13, @robonet_, @0xnonhuman [tweet 1/].

Method

Section titled “Method”

The tweet describes only the interface, not the internals. HandUMI is a wearable / robot-free variant of the UMI (Universal Manipulation Interface) recipe: a hand-held or hand-worn capture rig with matched grippers and wrist-camera-style observations that record the same modalities the eventual robot sees, so the demonstration format is embodiment-agnostic at collection time. To retarget to a new arm you physically swap the gripper module onto the target arm; the calibration + QA pass converts the recorded trajectories into a LeRobot v3 dataset that a policy training pipeline can consume. Two downstream execution modes are supported: simulator replay (for verification / synthetic augmentation) and real teleoperation. The linked video is not fetchable through the tweet page; details beyond this belong in a follow-up thread or repo.

Results

Section titled “Results”

None at filing — this is a product/tool launch tweet, not a paper. Views at capture: 2.8K. Engagement counts are the only numeric signals on the tweet itself.

Why it’s interesting

Section titled “Why it’s interesting”

HandUMI is a direct instantiation of the make-the-collection-tool-embodiment-invariant position on the Human-to-Robot Retargeting interface-location axis — the same shortcut Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training scaled to 100,000 hours for their XR-1 VLA — but packaged as an open-source, off-the-shelf hardware+software stack for small teams, and extended across four distinct bimanual arms rather than a single vendor rig. It complements UMI-Bench 1.0: An Open and Reproducible Real-World Benchmark for Tabletop Robotic Manipulation with UMI Data‘s UMI-data evaluation protocol by sitting upstream on the collection side, and contrasts with the algorithmic retargeting positions on that concept page (physics-in-loop ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting, reduced-DoF Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots, intent-level LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition, and full-articulation-at-scale EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data) by fixing the interface entirely in hardware.

See also

Section titled “See also”