HiFi-UMI-2K

Robot-free bimanual manipulation corpus captured with a handheld UMI rig: 482K demonstrations, 2,137 hours, six hardware-synchronised camera views per episode, shipped in LeRobot v3 format. No robot was involved in collection.

HiFi-UMI-2K is the public release from HiFi-UMI, a portable bimanual capture system built by Simple AI's Simple World Lab. Demonstrations are recorded by a human wearing a head-mounted stereo rig and a glove-like gripper on each hand — there is no robot, no teleoperation rig, and no instrumented room anywhere in the collection loop.

The claim being tested. Robot-free UMI data scales easily but is normally used only for pre-training, with a small amount of real-robot teleoperation added at post-training as an accuracy "anchor". The accompanying paper asks whether raising the fidelity of the robot-free data can remove that anchor entirely, and reports that a policy post-trained solely on HiFi-UMI deploys directly on a real bimanual robot — success-rate differences of −2.5, +3.1 and −0.6 points against in-domain teleoperation across StarVLA-QwenPI, OpenPI-π₀.₅ and LingBot-VA.

Capture. Hand pose comes from offline stereo-inertial SLAM plus fiducial marker cubes localised in the head-camera frame, giving roughly 3 mm workspace-local end-effector accuracy without external tracking. Both hands are solved in the same frame, so inter-hand relative pose is native rather than reconstructed across cameras. A shared GPIO trigger keeps every camera, IMU and gripper signal within 40 µs.

Six views per episode. A head-mounted stereo pair plus two non-parallel fisheye cameras on each hand, covering about 200° around each gripper — all six at 640×512, h264, 25 fps.

What is in the release. 482,060 episodes and 192,297,515 frames across 398 independently-loadable LeRobot v3 parts, totalling 16.0 TB. 588 distinct natural-language task strings, drawn from ordinary indoor work: microwaves and rice cookers, drawers and cabinets, whiteboards and meeting-room doors, wiping and tidying. `observation.state` and `action` are both float32[20] — per hand, `[x, y, z, rot6d ×6, gripper_angle_rad]` for right then left. The action is an absolute next-state target, not a relative delta, and the authors warn against assuming otherwise.

Validity masks, not success labels. Frames that fail quality control are deliberately retained so that Parquet rows, video frames and timestamps stay index-aligned; `valid.frame == true` is the intended training filter. There is no per-episode success or failure annotation in the exported metadata — the corpus is curated demonstrations that passed replay validation, so it is treated here as success-only.

The 2,000-hour public release is a curated subset of a larger internal corpus the paper describes as exceeding 20,000 hours. Released under CC BY 4.0.