Geometric deep learning Graph Neural Networks (GNNs) are a well-known kind of Deep Networks, allowing to enforce node-permutation equivariance (re-indexing), a symmetry present in all graphs. Likewise, rotation-equivariant networks allow to enforce rotation equivariance at the level of the very architecture. This is a game-changer that many are comparing to the introduction of CNNs and then GNNs, and is rapidly taking over the other methods for physics/chemistry- derived problems, where a learned representation of the input should not depend on the viewpoint (a rotation of the system of coordinates). Glassy materials: We propose a test case that offers ideal conditions for studying and designing new equivariant GNNs: glassy materials. The task is a node-wise regression on 3D point clouds. This test case matters in its own right, since understanding glassy dynamics is a long-standing open problem in physics. It is also a hard problem1 : solving it satisfactorily requires the design of extremely expressive and sensitive GNNs. It also differs from the more common benchmarks, which usually involve sparse assemblies of atoms (molecules, or at most an adsorbate on a surface): here, space is densely packed with particles. Predicting glassy dynamics has attracted a surge of interest recently [JAB+ 23], and several approaches now compete for the state of the art.