The speed/accuracy trade-off in Figure 1b is wild. Seeing EfficientFlow hit that ~53% success rate with just 1 Neural Function Evaluation (NFE=1) is a massive jump over EquiDiff, which looks painfully slow in comparison.
Usually, dropping to single-step generation kills performance in flow matching, but that "acceleration regularization" trick to straighten the trajectories seems to actually work. Getting inference down to 15ms while maintaining SE(3) equivariance is exactly what we need to move these policies out of simulation and onto real hardware where latency actually matters.