RationalVLA: A Rational Vision-Language-Action Model with Dual System

--A fundamental requirement for real-world robotic deployment is the ability to understand and respond to natural language instructions. Existing language-conditioned manipulation tasks typically assume that instructions are perfectly aligned with the environment. This assumption limits robustness and generalization in realistic scenarios where instructions may be ambiguous, irrelevant, or infeasible. T o address this problem, we introduce RAtional MAnipulation (RAMA), a new benchmark that challenges models with both unseen executable instructions and defective ones that should be rejected. In RAMA, we construct a dataset with over 14,000 samples, including diverse defective instructions spanning six dimensions: visual, physical, semantic, motion, safety, and out-of-context. We further propose the Rational Vision-Language-Action model (RationalVLA). It is a dual system for robotic arms that integrates the high-level vision-language model with the low-level manipulation policy by introducing learnable latent space embeddings. This design enables RationalVLA to reason over instructions, reject infeasible commands, and execute manipulation effectively. Experiments demonstrate that RationalVLA outperforms state-of-the-art baselines on RAMA by a 14.5% higher success rate and 0.94 average task length, while maintaining competitive performance on standard manipulation tasks. "Half of the troubles of this life can be traced to saying yes too quickly and not saying no soon enough. " -- Josh Billings Embodied intelligence represents the ultimate manifestation of artificial intelligence [1]. A necessary condition for the successful deployment of embodied intelligence in the real world is its ability to understand natural language and respond appropriately, either by providing correct answers or by executing the corresponding actions. This demand has sparked research on language-conditioned manipulation tasks, which require robots to follow natural language instructions to complete specific manipulation actions. Wenxuan Song, Jiayi Chen, Wenxue Li, Xu He, Xinhu Zheng, and Haoang Li are with The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China.