This work investigates whether continued pre‑training (CPT) can boost embodied learning in virtual worlds that differ in visual style, dynamics, controls, and objectives, as it does in physical environments. We conduct a controlled study across three axes—data volume, data diversity, and model size—creating 120 training configurations. Model sizes range from 0.8 B to 27 B parameters, data budgets span 50 to 1,600 hours, and data mixtures vary from Minecraft‑only to diverse multi‑world corpora. Performance and generalization are measured with three‑level validation losses and gameplay benchmarks covering both in‑corpus worlds and unseen worlds. The findings are:
-
More data improves in‑corpus performance; diversity aids transfer to related worlds but yields limited gains for distant worlds; larger models increase data efficiency but, when trained on a single world, generalize worse to distant worlds.
-
Training exhibits an early “interface phase” with broad cross‑world benefits, followed by a later “world phase” where gains are confined to in‑corpus and related worlds. An empirical scaling law is fitted, relating validation loss to data volume, diversity, and model size.
-
Our models are competitive with Minecraft specialists and state‑of‑the‑art generalist models on 3D benchmarks, and they excel in real‑time settings. Nevertheless, the ceiling of imitation learning and persistent gaps in distant worlds indicate that building truly generalist embodied agents requires richer training stages beyond CPT.
Review