Vision-Language-Action Models and World Models for Robotic Manipulation

Recent advances in vision-language-action (VLA) models are transforming robotic manipulation by enabling robots to connect visual observations and natural-language instructions directly with physical actions. Large-scale multimodal pretraining has significantly improved semantic understanding, instruction following, and action generation across a wide range of manipulation tasks. However, many existing VLA systems remain largely reactive, predicting actions from current observations without explicitly representing how objects, environments, and robot states may evolve. This limitation becomes particularly important in long-horizon, contact-rich, and dynamically changing tasks, where successful execution requires anticipation, continuous state estimation, and timely behavioral adjustment.World models offer a complementary approach by learning predictive representations of future observations, physical states, object interactions, and task progress. Their integration with VLA policies may allow robots to anticipate action consequences, compare possible outcomes, support task-level planning, identify execution deviations or failures, and adapt their actions through closed-loop interaction. Despite this potential, important challenges remain, including physically consistent multimodal prediction, effective interaction between policy and world-model components, uncertainty estimation, long-horizon reasoning, computational efficiency, and real-time deployment on physical robotic platforms.This Research Topic aims to bring together recent theoretical, methodological, and system-level advances in VLA models and world models for intelligent, adaptive, and reliable robotic manipulation. We particularly encourage studies investigating how predictive representations of the physical world can improve robotic perception, reasoning, planning, action generation, and closed-loop execution in realistic environments.Topics of interest include, but are not limited to:Vision-language-action foundation models for robotic manipulation;World models for predicting visual, geometric, tactile, and proprioceptive states;Joint, modular, or hierarchical integration of VLA policies and world models;Predictive representation learning for robot–object and robot–environment interactions;Contact-rich, dexterous, bimanual, mobile, and humanoid manipulation;Language-conditioned task planning, action reasoning, and long-horizon execution;Multimodal learning from robot demonstrations, videos, simulations, and interaction data;Self-supervised, imitation, reinforcement, and model-based learning for manipulation;Uncertainty-aware prediction, failure detection, recovery, and safety;Closed-loop planning and control using predicted future states;Sim-to-real transfer and scalable robotic data collection;Efficient architectures and real-time deployment on physical robots;Benchmarks, datasets, evaluation protocols, and reproducible robotic systems.By connecting embodied foundation models with predictive representations of the physical world, this Research Topic seeks to advance robotic systems that can understand instructions, anticipate future outcomes, and execute complex manipulation tasks with greater autonomy, robustness, and reliability. Topic editors: Moyun Liu, Xin Zhang, JIAWEI DU, Mobarak I. Hoque Source: https://www.frontiersin.org/research-topics/85292/vision-language-action-models-and-world-models-for-robotic-manipulation