Determine whether predicted representations add motion information beyond current representations

Determine whether the predicted optical-flow-supervised representation adds linearly decodable future-motion information beyond the current representation.

Background

Appendix D evaluates whether the current representation and the predicted representation encode future end-effector displacement. Frozen representations are assessed using a linear ridge-regression decoder trained on simulator-derived displacement labels over a 15-frame horizon. Although optical-flow supervision improves linear decoding relative to future-frame supervision, and the current representation outperforms the predicted representation, the reported comparison does not resolve whether the predictor contributes additional linearly decodable motion information beyond what is already present in the current representation.