FLAIR · project
ReorientDiff
A robot cannot always place an object at its target pose with a single pick-and-place action. It may first need to set the object down in an intermediate orientation that permits a different grasp. ReorientDiff learns to propose these intermediate poses from visual observations and a language description of the task.
Generating feasible intermediate poses
Scene information and the task prompt are combined into a joint representation that conditions a diffusion model. Classifier-free guidance steers pose generation, and gradients from learned feasibility scores refine the candidates. The model represents a distribution of possible reorientations, allowing several ways of completing a task to be considered.
Object manipulation experiments
The paper evaluates reorientation in simulation using YCB objects and a suction gripper. These experiments investigate whether learned pose proposals enable the required placement through intermediate configurations. The contribution is a conditional generative approach to reorientation planning, linking semantic task information with geometric manipulation constraints.
Publication
ReorientDiff: Diffusion Model based Reorientation for Object Manipulation
Utkarsh A. Mishra and Yongxin Chen
ICRA 2024