ObjDream: Guiding Vision-Language-Action Models with Object-Level Understanding and Future State Prediction

This paper presents ObjDream, an object-centric VLA framework that improves robotic manipulation by incorporating object-level understanding and future state prediction. The method uses a spatio-temporal object encoder to represent persistent object identity and dynamic state, which are integrated into the vision-language model. Action-conditioned object dynamics modules are used to supervise object state evolution during training, leading to better generalization in diverse scenarios.