ICP-Flow is a learning-free method for estimating scene flow from LiDAR scans captured by autonomous vehicles. The method leverages the assumption that objects in the scene move rigidly over time, using the Iterative Closest Point (ICP) algorithm to align clusters of points and estimate rigid transformations. A histogram-based initialization technique is proposed to improve ICP's performance by providing a good starting point for alignment. The complete scene flow is recovered from these transformations. ICP-Flow outperforms state-of-the-art baselines, including supervised models, on the Waymo dataset and performs competitively on Argoverse-v2 and nuScenes. Additionally, a feedforward neural network trained with pseudo labels generated by ICP-Flow achieves real-time inference with only marginal performance loss. The method is further extended to handle longer temporal gaps, up to 0.4 seconds, where other models fail. The contributions of ICP-Flow include a learning-free approach, incorporation of multi-body rigid-motion assumptions, and high-quality pseudo labels for real-time inference.ICP-Flow is a learning-free method for estimating scene flow from LiDAR scans captured by autonomous vehicles. The method leverages the assumption that objects in the scene move rigidly over time, using the Iterative Closest Point (ICP) algorithm to align clusters of points and estimate rigid transformations. A histogram-based initialization technique is proposed to improve ICP's performance by providing a good starting point for alignment. The complete scene flow is recovered from these transformations. ICP-Flow outperforms state-of-the-art baselines, including supervised models, on the Waymo dataset and performs competitively on Argoverse-v2 and nuScenes. Additionally, a feedforward neural network trained with pseudo labels generated by ICP-Flow achieves real-time inference with only marginal performance loss. The method is further extended to handle longer temporal gaps, up to 0.4 seconds, where other models fail. The contributions of ICP-Flow include a learning-free approach, incorporation of multi-body rigid-motion assumptions, and high-quality pseudo labels for real-time inference.