ICP-Flow is a learning-free method for estimating LiDAR scene flow, which represents the 3D motion between two consecutive LiDAR scans. The method leverages the assumption that objects in autonomous driving scenarios move rigidly, and uses the Iterative Closest Point (ICP) algorithm to align objects over time and estimate rigid transformations. A histogram-based initialization is proposed to provide a good starting point for ICP, improving the accuracy of the scene flow estimation. The method outperforms state-of-the-art baselines on the Waymo dataset and performs competitively on Argoverse-v2 and nuScenes. Additionally, a feedforward neural network is trained using pseudo labels generated by ICP-Flow, achieving real-time inference with minimal performance loss. The method is also extended to estimate scene flow over longer temporal gaps, up to 0.4 seconds, where other models fail. ICP-Flow provides high-quality pseudo labels for training neural networks and is capable of real-time inference. The method is validated on multiple datasets and shows competitive performance in scene flow estimation, particularly for dynamic foreground objects. The design is based on motion rigidity and does not require manual annotation or training data. The method is efficient and can handle large volumes of data, making it suitable for autonomous driving applications.ICP-Flow is a learning-free method for estimating LiDAR scene flow, which represents the 3D motion between two consecutive LiDAR scans. The method leverages the assumption that objects in autonomous driving scenarios move rigidly, and uses the Iterative Closest Point (ICP) algorithm to align objects over time and estimate rigid transformations. A histogram-based initialization is proposed to provide a good starting point for ICP, improving the accuracy of the scene flow estimation. The method outperforms state-of-the-art baselines on the Waymo dataset and performs competitively on Argoverse-v2 and nuScenes. Additionally, a feedforward neural network is trained using pseudo labels generated by ICP-Flow, achieving real-time inference with minimal performance loss. The method is also extended to estimate scene flow over longer temporal gaps, up to 0.4 seconds, where other models fail. ICP-Flow provides high-quality pseudo labels for training neural networks and is capable of real-time inference. The method is validated on multiple datasets and shows competitive performance in scene flow estimation, particularly for dynamic foreground objects. The design is based on motion rigidity and does not require manual annotation or training data. The method is efficient and can handle large volumes of data, making it suitable for autonomous driving applications.