Performance of a single multi-objective model across all modes

Show that a single multi-objective OpenWAM model jointly trained for policy, inverse-dynamics, and forward-dynamics objectives can match the strongest specialized OpenWAM variants across all operating modes.

Background

Appendix E evaluates a unified OpenWAM checkpoint trained jointly on policy, local-context inverse-dynamics, and local-context forward-dynamics objectives. Although the unified model outperforms the external UVA baseline on several metrics, its counterfactual dynamics performance is weaker than that of dedicated specialists and its policy score is lower than policy-specialized OpenWAM models. The authors therefore leave unresolved whether one multi-objective checkpoint can equal the best specialized models across all modes.

References

Appendix~E shows that policy, IDM, and FDM objectives can be trained jointly in one checkpoint, but we have not yet shown that a single multi-objective model can match the strongest specialized OpenWAM variants across all modes.

— OpenWAM: An Open Framework for Composable World-Action Models  (2610.07922 - Yu et al., 6 Oct 2026) in Discussion and Conclusion, Section 6; Appendix E, Section Unified Policy and Dynamics Modeling