Vision-Language Models (VLMs) can propose multi-step manipulation plans, but they have no grounded model of how a scene will change, and the pixel-level world models that could check those plans are slow and drift over long horizons. We present Verify2Act, a modular verification framework that separates semantic action proposal from latent physical verification. A VLM proposes candidate plans, a latent world model (V2A-WM) imagines each of them in the DINOv2 feature space, and a dual-head critic checks the imagined rollouts for temporal consistency and goal satisfaction before anything is executed. V2A-WM extends recent flow-matching world models with cross-attention action grounding and a causal temporal history window. We deploy the full system on a low-cost robot arm, training the world model and critic only on data generated in a digital twin. Across six language-conditioned tasks, Verify2Act reaches 83.3\% task success, compared with 61.7\% for a latent world model without our architectural changes, 53.3\% for the VLM alone, and 46.7\% for a pixel-diffusion world model. It verified every plan it executed and incurred no planning-induced failures: all of its failed episodes were caused by low-level grasp or placement slips. In simulation, Verify2Act obtains the highest five-task chain success on CALVIN among the compared baselines with 30$\times$ fewer reflective VLM calls than pixel-diffusion verification, and the highest success on cluttered nut assembly.
Sequential manipulation in clutter requires selecting which objects to remove and determining how to manipulate them. Learning both decisions jointly can complicate training and make failures difficult to diagnose. We present Unveiler, a modular framework that separates scene reasoning and object selection from action execution through a discrete object-index interface. Its Spatial Relationship Encoder (SRE) selects the best object to grasp, while a separate Action Decoder generates orientation-discretized push-grasp actions. The SRE is trained in two stages: imitation learning from heuristic demonstrations, followed by search-based policy improvement using simulated object removals. This training procedure requires no real-robot training data and allows the SRE to be paired with a different execution policy. We evaluate Unveiler in PyBullet simulation and on a physical Dofbot-Pro. In real-robot experiments, it retrieves two occluded targets in the prescribed order in 90\% of scenes, compared with 20-30\% for the heuristic, GPT-4o, and the imitation-only variant. In PyBullet, Unveiler achieves the highest task-completion rate across all six clutter-density and occlusion conditions, reaching 90.0\% under full occlusion with 6-9 objects, compared with 67.5\% for the heuristic and 60.0\% for ThinkGrasp. These results demonstrate the potential of modular, object-centric spatial reasoning for sequential manipulation in clutter.