Learning the RoPEs: Better 2D and 3D Position Encodings with STRING

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Schenck, Connor, Reid, Isaac, Jacob, Mithun George, Bewley, Alex, Ainslie, Joshua, Rendleman, David, Jain, Deepali, Sharma, Mohit, Dubey, Avinava, Wahid, Ayzaan, Singh, Sumeet, Wagner, René, Ding, Tianli, Fu, Chuyuan, Byravan, Arunkumar, Varley, Jake, Gritsenko, Alexey, Minderer, Matthias, Kalashnikov, Dmitry, Tompson, Jonathan, Sindhwani, Vikas, Choromanski, Krzysztof
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912221218471936
author Schenck, Connor
Reid, Isaac
Jacob, Mithun George
Bewley, Alex
Ainslie, Joshua
Rendleman, David
Jain, Deepali
Sharma, Mohit
Dubey, Avinava
Wahid, Ayzaan
Singh, Sumeet
Wagner, René
Ding, Tianli
Fu, Chuyuan
Byravan, Arunkumar
Varley, Jake
Gritsenko, Alexey
Minderer, Matthias
Kalashnikov, Dmitry
Tompson, Jonathan
Sindhwani, Vikas
Choromanski, Krzysztof
author_facet Schenck, Connor
Reid, Isaac
Jacob, Mithun George
Bewley, Alex
Ainslie, Joshua
Rendleman, David
Jain, Deepali
Sharma, Mohit
Dubey, Avinava
Wahid, Ayzaan
Singh, Sumeet
Wagner, René
Ding, Tianli
Fu, Chuyuan
Byravan, Arunkumar
Varley, Jake
Gritsenko, Alexey
Minderer, Matthias
Kalashnikov, Dmitry
Tompson, Jonathan
Sindhwani, Vikas
Choromanski, Krzysztof
contents We introduce STRING: Separable Translationally Invariant Position Encodings. STRING extends Rotary Position Encodings, a recently proposed and widely used algorithm in large language models, via a unifying theoretical framework. Importantly, STRING still provides exact translation invariance, including token coordinates of arbitrary dimensionality, whilst maintaining a low computational footprint. These properties are especially important in robotics, where efficient 3D token representation is key. We integrate STRING into Vision Transformers with RGB(-D) inputs (color plus optional depth), showing substantial gains, e.g. in open-vocabulary object detection and for robotics controllers. We complement our experiments with a rigorous mathematical analysis, proving the universality of our methods.
format Preprint
id arxiv_https___arxiv_org_abs_2502_02562
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Learning the RoPEs: Better 2D and 3D Position Encodings with STRING
Schenck, Connor
Reid, Isaac
Jacob, Mithun George
Bewley, Alex
Ainslie, Joshua
Rendleman, David
Jain, Deepali
Sharma, Mohit
Dubey, Avinava
Wahid, Ayzaan
Singh, Sumeet
Wagner, René
Ding, Tianli
Fu, Chuyuan
Byravan, Arunkumar
Varley, Jake
Gritsenko, Alexey
Minderer, Matthias
Kalashnikov, Dmitry
Tompson, Jonathan
Sindhwani, Vikas
Choromanski, Krzysztof
Machine Learning
Artificial Intelligence
Computer Vision and Pattern Recognition
Robotics
We introduce STRING: Separable Translationally Invariant Position Encodings. STRING extends Rotary Position Encodings, a recently proposed and widely used algorithm in large language models, via a unifying theoretical framework. Importantly, STRING still provides exact translation invariance, including token coordinates of arbitrary dimensionality, whilst maintaining a low computational footprint. These properties are especially important in robotics, where efficient 3D token representation is key. We integrate STRING into Vision Transformers with RGB(-D) inputs (color plus optional depth), showing substantial gains, e.g. in open-vocabulary object detection and for robotics controllers. We complement our experiments with a rigorous mathematical analysis, proving the universality of our methods.
title Learning the RoPEs: Better 2D and 3D Position Encodings with STRING
topic Machine Learning
Artificial Intelligence
Computer Vision and Pattern Recognition
Robotics
url https://arxiv.org/abs/2502.02562