MimicIK: Real-Time Generative Inverse Kinematics from Teleoperation with FK Consistency

MimicIK is a real-time generative inverse kinematics framework that utilizes conditional flow matching to learn robust joint-space motion priors from teleoperation data. It achieves a mean position error of 4.65 mm and a 92.01% success rate on a dataset of 8,848 demonstrations, significantly improving spatial accuracy and motion smoothness while reducing inference latency from 21.66 ms to 6.74 ms. The introduction of a differentiable forward-kinematics consistency loss allows for stable performance near singular configurations, making it suitable for real-time control in robotic applications.