Foundation model

Unified Inertial Encoder (Amazon FAR)

human + humanoid IMU encoder with motion–text alignment

Dataset: Nymeria
Year: 2026
Form factor: Multi site body IMU (12 sites) + humanoid robot

IMU arXiv 2611.10137Open codeOpen weights

Code Weights Cite

Chen, Yutian et al., "A Unified Inertial Encoder for Human and Humanoid Motion", arXiv:2611.10137, CC BY 4.0

PreprintCatalog paper date: Nov 6, 2026Added to atlas: Nov 10, 2026Source

GitHub starsView on GitHub HF downloadsView on Hugging Face

Activity counts measure attention, not model quality. Missing counts are unknown. Publication status follows the linked source; a listed venue alone does not confirm peer review.

55.6M-parameter inertial encoder trained on human motion with synthesized IMU and language plus retargeted humanoid motion; zero-shot 0.099 m/s velocity error on a real Unitree G1. Apache-2.0 code; ungated HF checkpoint with an unspecified ('other') weights licence.

Signals: IMU
Form factor: multi-site body IMU (12 sites) + humanoid robot
License: Apache-2.0 (code); weights 'other' (unspecified)
Year: 2026

Pretraining data: Nymeria (Project Aria), Virginia Tech Natural Motion, BONES-Seed (SOMA + G1 retargeted); eval also on real G1 (Open-IO/humanoid-g1-odometry)

At a glance

Type: Foundation model
Modalities: IMU
Form factor: Multi site body IMU (12 sites) + humanoid robot
Parameters: Not reported
Architecture: Not reported
Venue: arXiv 2611.10137 (2026-11-06)
Year: 2026
arXiv: 2611.10137 (opens in a new tab)
DOI: Not reported
Code: Open (opens in a new tab)
Weights: Open (opens in a new tab)
Licence: Apache-2.0 (code); weights 'other' (unspecified)
Caveats: Not reported
Datasets: Nymeria
Papers: A Unified Inertial Encoder for Human and Humanoid Motion
Added: Nov 10, 2026
Updated: Nov 10, 2026

Abstract

Learning humanoid state estimation and language-grounded motion understanding from robot recordings ties training coverage to the robot's motion repertoire. Human motion datasets offer broader coverage, yet transferring this knowledge requires overcoming differences in body geometry, sensing, and physical disturbances. To enable this transfer, we propose a unified inertial encoder whose shared inertial interface uses kinematic IMU synthesis to represent human and humanoid motion at corresponding body sites. Within this common input space, joint training on human and kinematically retargeted humanoid motion combines velocity estimation and motion–text alignment with continuous-time encoding and sensor and motion randomization. The resulting model transfers zero-shot to a physical Unitree G1 without physics simulation, real-robot training trajectories, or fine-tuning, achieving the lowest velocity error among evaluated methods across five datasets, including 0.099m/s on real G1. Its motion-description retrieval also extends to retargeted G1 motion categories observed only in human training, demonstrating how human motion knowledge can support the generalization of humanoid inertial odometry and language-grounded motion understanding.

Verbatim fromarXiv· arXiv metadata (abstracts are released under CC0)

How it works

Datasets linked in the hub; pretraining vs evaluation roles follow the paper.

Pretraining data: Nymeria (Project Aria), Virginia Tech Natural Motion, BONES-Seed (SOMA + G1 retargeted); eval also on real G1 (Open-IO/humanoid-g1-odometry)
Inputs: IMU
Approach: imu · synthetic imu · motion text alignment · continuous time tokenizer · inertial odometry · humanoid

Datasets used· 1

Nymeriaegocentric multimodal daily-motion dataset (Project Aria glasses + two miniAria wristbands + XSens full-body mocap; 264 participants, CC BY-NC 4.0)Pretraining264 subjects · IMU · Barometer · Audio

Papers· 1

A Unified Inertial Encoder for Human and Humanoid MotionarXiv preprint · 2026arXiv preprint · 2026

Cite this model

Cite the publication recorded for this model. Check it against the publisher before submitting.

Chen, Y., Zhao, S., Duan, R., Liu, K., Sferrazza, C., Shi, G., Qiu, Y., & Sreenath, K. (2026). A Unified Inertial Encoder for Human and Humanoid Motion. arXiv 2611.10137. https://arxiv.org/abs/2611.10137

Chen, Yutian, Shibo Zhao, Rocky Duan, Karen Liu, Carmelo Sferrazza, Guanya Shi, Yuheng Qiu, and Koushil Sreenath. 2026. "A Unified Inertial Encoder for Human and Humanoid Motion." arXiv 2611.10137. https://arxiv.org/abs/2611.10137.

Chen, Y., Zhao, S., Duan, R., Liu, K., Sferrazza, C., Shi, G., Qiu, Y. and Sreenath, K. (2026) 'A Unified Inertial Encoder for Human and Humanoid Motion', arXiv 2611.10137. Available at: https://arxiv.org/abs/2611.10137

Chen, Yutian, et al. "A Unified Inertial Encoder for Human and Humanoid Motion." arXiv 2611.10137, 2026. https://arxiv.org/abs/2611.10137.

Y. Chen et al., "A Unified Inertial Encoder for Human and Humanoid Motion," arXiv 2611.10137, 2026, [Online]. Available: https://arxiv.org/abs/2611.10137.

@misc{chen2026unified,
  title         = {{A Unified Inertial Encoder for Human and Humanoid Motion}},
  author        = {Chen, Yutian and Zhao, Shibo and Duan, Rocky and Liu, Karen and Sferrazza, Carmelo and Shi, Guanya and Qiu, Yuheng and Sreenath, Koushil},
  year          = {2026},
  howpublished  = {arXiv 2611.10137},
  eprint        = {2611.10137},
  archivePrefix = {arXiv}
}