VTLA: Vision-Tactile-Language-Action Model with Preference Learning for Insertion Manipulation

2025-05-14

AI Summary

This work addresses the challenge of language-conditioned, contact-intensive robotic manipulation (e.g., fingertip insertion), focusing on robust fusion of vision–tactile multimodal signals for language-guided policy learning. We propose the first unified vision–tactile–language–action modeling framework, enabling deep cross-modal perception integration via language alignment. Crucially, we introduce a novel Direct Preference Optimization (DPO) paradigm tailored for continuous control—replacing conventional token-level classification losses. To support training, we construct a low-cost, simulation-generated multimodal instruction dataset (vision–tactile–action–instruction). Experiments demonstrate over 90% success rate on unseen peg-insertion tasks, substantially outperforming diffusion-based policies and TLA/VLA baselines. Moreover, our method exhibits strong sim-to-real transfer capability.

Technology Category

Intelligent Robots: Multimodal Perception & Sensor FusionComputer Vision: Language and VisionNatural Language Processing: Language Grounding & Multi-modal NLP

Application Category

Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web searchUser Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendationSemantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

While vision-language models have advanced significantly, their application in language-conditioned robotic manipulation is still underexplored, especially for contact-rich tasks that extend beyond visually dominant pick-and-place scenarios. To bridge this gap, we introduce Vision-Tactile-Language-Action model, a novel framework that enables robust policy generation in contact-intensive scenarios by effectively integrating visual and tactile inputs through cross-modal language grounding. A low-cost, multi-modal dataset has been constructed in a simulation environment, containing vision-tactile-action-instruction pairs specifically designed for the fingertip insertion task. Furthermore, we introduce Direct Preference Optimization (DPO) to offer regression-like supervision for the VTLA model, effectively bridging the gap between classification-based next token prediction loss and continuous robotic tasks. Experimental results show that the VTLA model outperforms traditional imitation learning methods (e.g., diffusion policies) and existing multi-modal baselines (TLA/VLA), achieving over 90% success rates on unseen peg shapes. Finally, we conduct real-world peg-in-hole experiments to demonstrate the exceptional Sim2Real performance of the proposed VTLA model. For supplementary videos and results, please visit our project website: https://sites.google.com/view/vtla

Problem

• Bridging vision-language models for contact-rich robotic manipulation tasks
• Integrating visual-tactile inputs via cross-modal language grounding
• Enhancing Sim2Real performance in insertion tasks with preference learning

Innovation

• Integrates visual and tactile inputs via language grounding
• Uses Direct Preference Optimization for continuous tasks
• Achieves high success on unseen shapes with VTLA