RoboAlign: Learning Test-Time Reasoning for Language-Action Alignment in Vision-Language-Action Models

2026-03-22

AI Summary

This work addresses the stability and efficiency bottlenecks in existing vision–language–action models when translating the embodied reasoning capabilities of multimodal large language models into low-level actions. The authors propose RoboAlign, a novel framework that, for the first time, integrates zero-shot natural language reasoning at test time with reinforcement learning to achieve reliable alignment between language instructions and low-level actions using less than 1% of the training data. By combining supervised fine-tuning with a diffusion-based action head, RoboAlign significantly improves knowledge transfer from multimodal foundation models to action modules, outperforming SFT baselines by 17.5%, 18.9%, and 106.6% on the LIBERO and CALVIN benchmarks and in real-world robotic environments, respectively.

Technology Category

• Computer Vision: Language and Vision
• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Improving embodied reasoning in multimodal-large-language models (MLLMs) is essential for building vision-language-action models (VLAs) on top of them to readily translate multimodal understanding into low-level actions. Accordingly, recent work has explored enhancing embodied reasoning in MLLMs through supervision of vision-question-answering type. However, these approaches have been reported to result in unstable VLA performance, often yielding only marginal or even negative gains. In this paper, we propose a more systematic MLLM training framework RoboAlign that reliably improves VLA performance. Our key idea is to sample action tokens via zero-shot natural language reasoning and refines this reasoning using reinforcement learning (RL) to improve action accuracy. As a result, RoboAlign bridges the modality gap between language and low-level actions in MLLMs, and facilitate knowledge transfer from MLLM to VLA. To validate the effectiveness of RoboAlign, we train VLAs by adding a diffusion-based action head on top of an MLLM backbone and evaluate them on major robotics benchmarks. Remarkably, by performing RL-based alignment after SFT using less than 1\% of the data, RoboAlign achieves performance improvements of 17.5\%, 18.9\%, and 106.6\% over SFT baselines on LIBERO, CALVIN, and real-world environments, respectively.

Problem

• embodied reasoning
• vision-language-action models
• multimodal-large-language models
• language-action alignment
• modality gap

Innovation

• RoboAlign
• reinforcement learning
• vision-language-action alignment
• multimodal large language models
• test-time reasoning