The field of robot manipulation is witnessing significant advancements with the integration of novel frameworks and techniques. Researchers are exploring innovative methods to enhance spatial understanding, improve physical reasoning, and develop more efficient policy models. One of the key directions is the incorporation of equivariant models, which have been shown to improve data efficiency in diffusion policy learning. Another area of focus is the development of generative frameworks that can flexibly integrate both equality and inequality constraints, enabling more precise and adaptive constraint handling. These advancements have the potential to significantly improve the performance and generalizability of robot manipulation tasks. Notable papers include: Adaptive Diffusion Constrained Sampling, which proposes a generative framework for flexibly integrating constraints into an energy-based diffusion model. Scan, Materialize, Simulate, which presents a unified framework for physically grounded robot planning that combines 3D scene reconstruction, semantic segmentation, and physics simulation. SEM, which introduces a novel diffusion-based policy framework that enhances spatial understanding through spatial and embodiment-aware representations. 3D Equivariant Visuomotor Policy Learning via Spherical Projection, which enables SO(3)-equivariant policy learning using only monocular RGB inputs.