GraphDiff-IK: Graph diffusion solves whole-body inverse kinematics for any robot

Inverse kinematics (IK) remains a fundamental challenge in robotics, requiring the generation of joint configurations that satisfy target end-effector poses. Existing methods often fail to generalize across diverse robot morphologies or handle the multi-modal nature of IK, especially for articulated systems with multiple kinematic branches (e.g., dual-arm robots with torsos). Researchers from the team behind GraphDiff-IK introduce a novel framework that leverages graph diffusion to directly generate joint configurations on a kinematic graph representation of the robot.

The method constructs a graph from the robot's URDF, where nodes are actuated joints and edges encode kinematic dependencies. It then formulates IK as a conditional graph diffusion process, adding hierarchical stage-wise message passing and torso-aware conditioning for multi-branch robots. To ensure geometric consistency, they incorporate noisy forward kinematics feedback and task-space supervision during denoising. Extensive experiments across various robotic platforms demonstrate that GraphDiff-IK produces accurate, stable IK solutions while preserving the ability to generate multiple feasible configurations for redundant systems.