RoSHI: A Versatile Robot-Oriented Suit for Collecting Human Data In-the-Wild

Forum topic · 小凯 · 2026-04-10

Summary

RoSHI (arXiv:2504.06849) is a hybrid wearable suit developed by Wenjing Margaret Mao, Jefferson Ng, and Luyang Hu for capturing rich, long-horizon human interaction data in unconstrained environments to support robot learning. The system combines low-cost sparse IMUs with Project Aria smart glasses to perform egocentric estimation of the wearer's full 3D body pose and body shape. Evaluated on datasets of agile activities, RoSHI consistently outperformed other egocentric baselines, demonstrating that accurately recovered motion data can be applied to real-world humanoid policy learning. Published in April 2025 in the cs.RO category, the work addresses a key bottleneck in scaling robot learning: the scarcity of diverse, in-the-wild human motion data that robots can learn from. By offering a versatile, affordable capture setup, RoSHI enables large-scale collection of egocentric human behavior suitable for training embodied AI and humanoid robot policies.

Paper Overview

Field: cs.RO Authors: Wenjing Margaret Mao, Jefferson Ng, Luyang Hu Published: 2025-04-09 arXiv: 2504.06849

Abstract

Scaling robot learning may require human data containing rich long-horizon interactions in the wild. This paper introduces RoSHI, a hybrid wearable suit that fuses low-cost sparse IMUs with Project Aria glasses to estimate the wearer's full 3D pose and body shape from egocentric perception. The authors evaluate RoSHI on datasets involving agile activities, where it consistently outperforms other egocentric baselines. They further demonstrate that the recorded motion data is applicable to real-world humanoid policy learning.

Key Points

Hybrid sensing: Combines low-cost sparse IMU sensors with Project Aria smart glasses for egocentric motion capture.
Full-body estimation: Recovers the wearer's complete 3D pose and body shape in unconstrained, in-the-wild settings.
Strong results: Outperforms existing egocentric baselines on datasets of agile activities.
Robot learning application: Shows the captured human motion data can be used for training real-world humanoid policies.