The field of robotic manipulation and autonomous exploration is moving towards more robust and versatile methods, leveraging structured intermediate representations and diffusion-based policies to improve performance in complex tasks. Researchers are exploring the use of 3D flow, spectral rewards, and discrete diffusion policies to enable robots to better understand and interact with their environment. Notable developments include the integration of global graph inference with diffusion-based decision-making for autonomous exploration, and the introduction of spatial memory mechanisms in deep feature maps for 3D action policies.

Noteworthy papers include: 3D Flow Diffusion Policy, which achieves state-of-the-art performance in robotic manipulation tasks by leveraging scene-level 3D flow as a structured intermediate representation. Query-Centric Diffusion Policy, which improves performance in robotic assembly tasks by utilizing queries comprising objects, contact points, and skill information to guide low-level policies. TopoCut, which introduces a comprehensive benchmark for multi-step robotic cutting tasks and proposes an integrated policy learning pipeline for goal-conditioned policy learning. GUIDE, which synergistically combines global graph inference with diffusion-based decision-making for autonomous exploration, achieving up to 18.3% faster coverage completion and a 34.9% reduction in redundant movements. mindmap, which introduces a 3D diffusion policy that generates robot trajectories based on a semantic 3D reconstruction of the environment, showing effectiveness in solving tasks where state-of-the-art approaches without memory mechanisms struggle.