Contact-VLA: Zero-Shot Planning and Control for Contact-Rich Manipulation
Lists
Only you see this
Abstract
Vision-Language-Action (VLA) systems often lack adaptability and explainability due to their black-box structure and dependency on fixed action sets from extensive tele-operated datasets, limiting their effectiveness in complex, dynamic manipulation scenarios. To address this issue, we propose Contact-Rich Adaptive LLM-based Control (CoRAL), a novel modular framework capable of effectively managing complex, dynamic, and contact-rich manipulation tasks. By integrating foundational vision and language models with motion planning and reactive controllers, our system achieves zero-shot planning and adaptive manipulation without relying on extensive tele-operated action datasets. Unlike conventional VLAs, we explicitly separate the roles of vision models and Large Language Models (LLM): the vision module handles environmental parameter initialization and object pose tracking, while the LLM generates initial contact strategies and cost function estimations. This collaboration establishes a physical understanding of the scene, instantiated as a dynamic planning world model for our planner. Additionally, this modular approach significantly enhances both the explainability and performance of the overall framework, as demonstrated by ablation studies. Furthermore, we introduce a memory unit to leverage past manipulation experiences, enabling the generalization and efficient reuse of learned contact strategies and parameter adjustments across diverse manipulation scenarios. Experiments conducted on challenging contact-rich tasks validate our framework's robustness and highlight the critical design elements that contribute to its effectiveness.
BibTeX
@misc{cicek2026contact,
author = {Berk Cicek and Mert Kaan Er and Ege Doganay and Ozgur S. Oguz},
title = {{Contact-VLA: Zero-Shot Planning and Control for Contact-Rich Manipulation}},
year = {2026},
url = {https://openreview.net/forum?id=qPbDM5L8tE},
howpublished = {OpenReview},
note = {Submitted to ICLR 2026},
}