With the progress of 3D human pose and shape estimation, state-of-the-artmethods can either be robust to occlusions or obtain pixel-aligned accuracy innon-occlusion cases. However, they cannot obtain robustness and mesh-imagealignment at the same time. In this work, we present NIKI (Neural InverseKinematics with Invertible Neural Network), which models bi-directional errorsto improve the robustness to occlusions and obtain pixel-aligned accuracy. NIKIcan learn from both the forward and inverse processes with invertible networks.In the inverse process, the model separates the error from the plausible 3Dpose manifold for a robust 3D human pose estimation. In the forward process, weenforce the zero-error boundary conditions to improve the sensitivity toreliable joint positions for better mesh-image alignment. Furthermore, NIKIemulates the analytical inverse kinematics algorithms with the twist-and-swingdecomposition for better interpretability. Experiments on standard andocclusion-specific benchmarks demonstrate the effectiveness of NIKI, where weexhibit robust and well-aligned results simultaneously. Code is available athttps://github.com/Jeff-sjtu/NIKI