UMI-Style Glove Data Collection / 手套式数据采集

Definition

UMI-style glove data collection is a body-free manipulation-data method in which a person wears a sensor-rich handheld device or glove so that hand motion, camera views, timing, and sometimes pose or action information can be recorded as robot-trainable trajectories without operating a robot during collection.

Current Synthesis

The two bounded sources describe glove collection as a bridge between cheap but action-poor human video and precise but expensive real-robot teleoperation. 深普智能 emphasizes measurement quality and validation: its six-view device uses camera-based millimeter localization and microsecond synchronization without an external base station, and the team says trajectory replay reproduced more than 95% of collected trajectories on the body before it scaled collection. It later open-sourced about 2,000 hours, reported internal data in the tens of thousands of hours, and disclosed annotation and usability bottlenecks.

Liber AI emphasizes learning efficiency. Its source says natural glove motion aligns better with human video than slow, viewpoint-mismatched teleoperation and adds precise action information absent from ordinary internet video. The company reports that 10,000 glove hours outperformed two million internet-video hours, and that glove data combined with 25% of its real-robot teleoperation data beat a full teleoperation-only baseline. The synthesis is not that gloves eliminate real-robot data: they may reduce its share while robot deployment remains necessary for embodiment calibration, contact, evaluation, and final reliability.

Key Claims

• Glove collection can add precise human-action information that ordinary internet video lacks while avoiding the full cost and throughput limits of operating robots during every recording.
• Device precision, synchronization, occlusion handling, annotation, and trajectory replay determine whether nominal collection hours become usable robot data.
• Glove motion may align more naturally with human video than robot teleoperation, but cross-embodiment mapping is still required before execution.
• Real-robot data remains necessary for body calibration, contact dynamics, evaluation, and deployment even if glove data substitutes for part of the teleoperation set.
• Open data can function as research infrastructure and as a commercial lead channel, but download volume is not evidence of deployed robot performance.
• Collection-device usability and sensor choices can create heat, marker-size, tactile, and format-lock-in constraints.

Evidence

• Device precision and validation: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 describes the six views, millimeter localization, microsecond synchronization, removal of the base station, replay testing, and reported over-95% reproduction rate.
• Scale, annotation, and usability: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 gives the open and internal hours, download and purchase interest, annotation pipeline, marker-size issue, and glove-heat feedback.
• Human-to-robot bridge: 187: 对话LiberAI刘松铭：Astra来袭，00后物理AI创始人看到的胜率 explains the missing-action problem in internet video, the natural-motion argument against teleoperation, and embodiment conditioning.
• Reported substitution efficiency: 187: 对话LiberAI刘松铭：Astra来袭，00后物理AI创始人看到的胜率 reports the 10,000-hour versus two-million-hour comparison and the glove-plus-25%-teleoperation ablation.

Counterevidence & Qualifications

Device designs, replay rates, download counts, annotation quality, performance comparisons, substitution ratios, and cost estimates are company-reported and not independently verified. The sources do not show common tasks, bodies, protocols, or data-accounting rules that would make their hour counts directly comparable. The Shenpu open dataset has no tactile channel, so glove pipelines are not a complete physical-data strategy. Self-built collection hardware also creates a format commitment, and natural human motion can still differ from robot kinematics, force limits, sensing, and contact dynamics.

What Changed

• Added Liber AI’s use of glove data as an action-information bridge between internet video and robot control.
• Added the reported data-ablation case for partial substitution of real-robot teleoperation.
• Clarified that glove efficiency does not remove embodiment, contact, evaluation, or final-deployment requirements.

Related Concepts

• Embodied Robot Data Paradigms - collection-method landscape in which UMI-style gloves are one paradigm.
• Real Robot Data Strategy - broader data-recipe problem that self-built collection feeds.
• Robot Data Scale Up - scaling question that replay validation and cleaning address.
• Robot Teleoperation and Remote Takeover - adjacent high-quality collection method that operates the robot directly.
• Tactile Data Staging / 触觉数据暂缓采集 - decision to leave the glove-based dataset without tactile data for now.
• Egocentric Robot Data - first-person human data route that this glove-based method resembles but does not equal.
• World Action Models - action-conditioned model family that can consume mixed human, glove, and robot data.
• Liber AI - company reporting glove-data substitution and cross-embodiment results.