AIAny
Icon for item

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

Provides a portable, robot-free UMI capture pipeline and shows that policies post-trained only on this high-fidelity data deploy directly on real robots matching teleoperation baselines. Capture achieves ~3 mm end-effector accuracy, microsecond sync, ultra-wide FOV, and releases 2,000h HiFi-UMI-2K.

Introduction

Raising the fidelity of robot-free UMI demonstrations — rather than merely increasing the small fraction of real-robot anchors — can enable zero-robot post-training: policies trained or post-trained only on a sufficiently accurate UMI corpus can transfer to real hardware with minimal real-robot data.

Key Findings
  • Zero-robot post-training works in practice: a policy post-trained solely on the HiFi-UMI corpus deploys on a real robot and matches in-domain teleoperation baselines across three backbone families, with success-rate differences of -2.5, +3.1, and -0.6 percentage points.
  • High-fidelity capture matters: the HiFi-UMI system attains ~3 mm workspace-local end-effector accuracy without external trackers by combining head-mounted stereo-inertial SLAM, native inter-gripper relative pose, microsecond GPIO synchronization, and two wide-angle cameras per hand (~200° coverage).
  • Large-scale pretraining amplifies gains: pretraining on 4,000 hours from the same corpus reduced action error on ten unseen tasks by 41% and improved real-robot success on StarVLA-QwenPI by 18.1 percentage points.
  • Practical resource release: the authors open-source HiFi-UMI-2K, a 2,000-hour, microsecond-synchronized ultra-wide-FoV demonstration corpus with automatic reconstruction and simulation-validated replay.
Who it's for and trade-offs

Great fit if you develop robotic manipulation policies and need scalable, low-cost demonstration data that can transfer to real robots without large teleoperation budgets. It benefits teams focused on vision-language-action and world-action-model architectures seeking to reduce real-robot collection. Look elsewhere if your domain demands contact-rich dynamics measured by force/torque sensors, sub-millimeter positional accuracy, or specialized hardware not covered by head-mounted capture; HiFi-UMI emphasizes trajectory/pose fidelity and visual coverage rather than instrumented force sensing.

Where it fits

HiFi-UMI sits between small-scale teleoperation datasets (high fidelity but costly) and large, low-fidelity synthetic or proxy corpora: it aims to deliver near-teleop realism at scale, trading expensive robot-hours for a calibrated, synchronized human-mounted capture rig and large annotation-free hours of demonstrations.

Information

  • Websitearxiv.org
  • AuthorsYuteng Wei, Jinming Ma, Jiawei Wang, Weitao Zhou, Yushen Zuo, Ke Rui, Minglei Li, Jinhao Zhang, Zhikang Pan, Xiang Wang
  • Published date2026/07/28

More Items

Real-time streaming multimodal foundation model that uses a codec-native tokenizer (Mage-ViT) to encode motion- and residual-rich regions from video I/P frames, reducing visual token usage by over 75% and enabling up to ~3.5× wall-clock inference speedup after training on ~560M images and 100M video frames.

Recovers editable design files from raster images by growing an editable layer hierarchy via an agentic pipeline that selects and composes modality-specific tools. Introduces graceful verification (accept/prune/retry) to prevent error accumulation and presents the Figma Edit Replay Benchmark (909 files, 14,796 edits) to measure editability across layout, color, and text edits.

Drives long‑horizon desktop agents by reading and manipulating program state (files, DOM, backends) instead of relying on screenshots. The main agent uses code for actions and structural verification while a lightweight GUI subagent handles rare screenshot-click steps, improving success rates and lowering per-task cost versus screenshot-only approaches.