DeCo: Task Decomposition and Skill Composition for Zero-Shot Generalization in Long-Horizon 3D Manipulation
Zixuan Chen, Junhui Yin, Yangtao Chen, Jing Huo, Pinzhuo Tian, Jieqi Shi, Yiwen Hou, Yinchuan Li
Abstract
Generalizing language-conditioned multi-task imitation learning (IL) models to novel long-horizon 3D manipulation tasks is challenging. To address this, we propose <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DeCo</b> (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Task</i> <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">De</b><italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">composition and Skill</i> <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Co</b><italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mposition</i>), a model-agnostic framework that enhances zero-shot generalization to compositional long-horizon manipulation tasks. DeCo decomposes IL demonstrations into modular atomic tasks based on gripper-object interactions, creating a dataset that enables models to learn reusable skills. At inference, DeCo uses a vision-language model (VLM) to parse high-level instructions, retrieve relevant skills, and dynamically schedule their execution. A spatially-aware skill-chaining module ensures smooth, collision-free transitions between skills. We introduce <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DeCoBench</monospace>, a benchmark designed to evaluate compositional generalization in long-horizon manipulation tasks. DeCo improves the success rate of three IL models—RVT-2, 3DDA, and ARP—by <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">66.67%, 21.53%, and 57.92%</b>, respectively, on 12 novel tasks. In real-world experiments, the DeCo-enhanced model, trained on only 6 atomic tasks, completes 9 novel tasks in zero-shot, with a <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">53.33%</b> improvement over the baseline model. Project website: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://deco226.github.io</uri>.
BibTeX
@inproceedings{ral2026_decotaskdecompos,
title = {DeCo: Task Decomposition and Skill Composition for Zero-Shot Generalization in Long-Horizon 3D Manipulation},
author = {Zixuan Chen and Junhui Yin and Yangtao Chen and Jing Huo and Pinzhuo Tian and Jieqi Shi and Yiwen Hou and Yinchuan Li and Yang Gao},
booktitle = {RA-L 2026},
year = {2026}
}