HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone

2026-07-28

AI Summary

This work addresses the scarcity of high-fidelity, scalable real-world data for robotic manipulation policy learning by introducing HiFi-UMI, a system that enables training and direct deployment of policies on physical robots using only high-quality, robot-free UMI (Unconstrained Manipulation Interaction) data—without requiring teleoperated “anchor” demonstrations or fine-tuning. HiFi-UMI integrates head-mounted binocular-inertial SLAM, native gripper pose estimation, microsecond-level GPIO synchronization, and ultra-wide dual cameras (providing ~200° field of view per hand) to establish an end-to-end sim-to-real transfer paradigm. Evaluated on real robots, the resulting policies achieve up to 85% task success—comparable to human teleoperation—and, leveraging 4,000 hours of pretraining, reduce action error on unseen tasks by 41% and improve success rates by up to 18.1 percentage points. The HiFi-UMI-2K dataset is publicly released.

Abstract

Learning deployable manipulation policies is bottlenecked by the scarcity of data that is both high-fidelity and scalable. Real-robot teleoperation is accurate but costly to scale; robot-free UMI capture scales readily, and current practice uses the resulting data mainly for pre-training, adding a small real-robot "anchor" at post-training. We ask whether raising the fidelity of robot-free UMI data, rather than shrinking the real-robot fraction, can remove that anchor. We present HiFi-UMI, a portable UMI data-production system co-designed for trajectory accuracy, inter-gripper relative pose, synchronization, and field of view: head-mounted offline stereo-inertial SLAM, native rather than reconstructed relative pose, a shared microsecond GPIO trigger, and two wide-angle cameras per hand covering ~200 degrees. It reaches 3 mm workspace-local end-effector accuracy without external tracking infrastructure. Using this corpus, we demonstrate zero-robot post-training: a policy post-trained solely on HiFi-UMI demonstrations deploys directly on a real robot and matches in-domain teleoperation across three backbones spanning the vision-language-action and world-action-model families, with success-rate differences of -2.5, +3.1, and -0.6 percentage points on StarVLA-QwenPI, OpenPI-pi_0.5, and LingBot-VA; the strongest policy reaches 85% on a precision insertion task, even though the teleoperation baseline is collected in the evaluation scene and no HiFi-UMI trajectory is. Pre-training on 4,000 hours from the same corpus lowers action error on ten unseen tasks by 41% and, on StarVLA-QwenPI, raises real-robot success by a further 18.1 percentage points. We open-source HiFi-UMI-2K, 2,000 hours of microsecond-synchronized, ultra-wide-FoV demonstrations, each automatically reconstructed and validated through simulation replay, as a large-scale, high-fidelity resource for the robot-learning community.

Problem

• deployable manipulation policies
• high-fidelity data
• robot-free UMI
• real-robot teleoperation
• data scalability

Innovation

• HiFi-UMI
• zero-robot post-training
• high-fidelity demonstration
• stereo-inertial SLAM
• microsecond synchronization