Understanding the dynamics of slope instabilities is critical to mitigate the associated hazards, but their direct observation is often difficult due to their remote locations and their spontaneous nature. Seismology allows us to get unique information on these events, including on their dynamics. However, the link between the properties of these events (mass and kinematics) and the seismic signals generated is still poorly understood. We conducted a controlled rockfall experiment in the Riou Bourdoux torrent (southern French Alps) to try to better decipher those links. We deployed a dense seismic network and inferred the dynamics of the block from the reconstruction of the 3D trajectory from terrestrial and airborne high-resolution stereophotogrammetry. We propose a new approach based on machine learning to predict the mass and the velocity of each block. Our results show that we can predict those quantities with average errors of approximately 10 % for the velocity and 25 % for the mass. These accuracies are as good as or better than those obtained by other approaches, but our approach has the advantage in that it does not require the source to be localised, nor does it require a high-resolution velocity model or a strong assumption on the seismic wave attenuation model. Finally, the machine learning approach allows us to explore more widely the correlations between the features of the seismic signal generated by the rockfalls and their physical properties, and it might eventually lead to better constraints on the physical models in the future.