analiza aiqbench
This paper addresses a critical limitation in compositional diffusion policies for robotics by introducing a score-decorrelation loss that enforces specialization among task components. The method effectively reduces double-counting and improves interpretability without requiring significant architectural changes or computational overhead, advancing the state of modular multitask learning.
Learning to Be Different: Component Specialization for Compositional Diffusion Policies
arXiv:2610.27194v1 Announce Type: new Abstract: Diffusion policies with linearly composed scores have emerged as a promising approach to multitask robotic learning by combining separate factors to enable modular, task-wise adaptation, compositional generalization, and interpretable functional decomposition. However, current training objectives do not encourage factor specialization, often resulting in uninterpretable denoisers whose composed prediction is an average over identical copies, undermining their architectural premise. To address this, we propose a drop-in score-decorrelation auxiliary loss, motivated by projective composition theory. Our approach encourages specialization by penalizing overlap directly between component noise predictions, with little architectural change or computational overhead. We demonstrate across multiple benchmarks that our decorrelation loss consistently reduces measurable double-counting between components, improves task decomposition, and produces behaviorally meaningful specialization without compromising multitask performance. Website: https://al3cl1.github.io/robotics-score-decorrelation.
Abstract:Diffusion policies with linearly composed scores have emerged as a promising approach to multitask robotic learning by combining separate factors to enable modular, task-wise adaptation, compositional generalization, and interpretable functional decomposition. However, current training objectives do not encourage factor specialization, often resulting in uninterpretable denoisers whose composed prediction is an average over identical copies, undermining their architectural premise. To address this, we propose a drop-in score-decorrelation auxiliary loss, motivated by projective composition theory. Our approach encourages specialization by penalizing overlap directly between component noise predictions, with little architectural change or computational overhead. We demonstrate across multiple benchmarks that our decorrelation loss consistently reduces measurable double-counting between components, improves task decomposition, and produces behaviorally meaningful specialization without compromising multitask performance. Website: this https URL.
Robotics (cs.RO)
arXiv:2610.27194 [cs.RO]
(or arXiv:2610.27194v1 [cs.RO] for this version)
https://doi.org/10.48550/arXiv.2610.27194 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 216 cuvinte
Subiecte conexe
Proof that the Klein-Gordon type equation with alpha attractor potential has no Liouvillian solution or as a composition of special functions
arXiv · 8 iun. 2026
Carving Radiology Reports at Their Evidential Joints: Anatomical Region Sets as Generation Units
arXiv · 22 oct. 2026
Post-Training in Time Series Foundation Models: A Unifying Framework
arXiv · 23 iul. 2026
GFFMERGE: Efficient Merging of Graph Neural Force Fields and Beyond
arXiv · 3 iun. 2026