We compare MXT+ with and without cross-embodiment pretraining against ACT and π_0.5 on 13 task–embodiment pairs. Under ID evaluation, pretraining yields particularly large gains on G1-Dex5 Sort, LocoMan-Dex2 Scoop, and xArm7-Dex4 Scoop. The benefit becomes even more pronounced under OOD evaluation, with substantial improvements on G1-Dex5 Kitchen-Bread, G1-Dex5 Insert, and Z1-Tool Scoop.

Main Results. Cross-embodiment-pretrained MXT+ generally improves task progress score across 13 task–embodiment pairs, with substantially larger gains under OOD evaluation.

We vary the pretraining mixture: no pretraining, human-only, human plus other robots, and full cross-embodiment data, which also includes the target robot's demonstrations on other tasks. Every variant is fine-tuned on the same target-embodiment, target-task demonstrations.

Pretraining Data Composition. Every pretraining mixture improves average OOD performance over no pretraining, and the gains grow as more embodiments are added. Full cross-embodiment pretraining performs best, improving the average task progress score by 0.12 in-distribution and 0.23 out-of-distribution, with OOD gains on all six task–embodiment pairs.

Embodiment dreaming asks the shared trunk to predict future embodiment latents built from egocentric image features, robot joint configurations, and end-effector joint configurations. EMA copies of MXT+'s own modality encoders provide stable targets. The objective encourages embodiment-aware dynamics modeling during pretraining and fine-tuning; dreaming heads and EMA teachers are discarded at deployment.

We compare MXT+ with MXT, its counterpart without the dreaming objective, on seven task–embodiment settings across five embodiments. The gains hold across tasks and embodiments, and are particularly pronounced on G1-Dex5.

Embodiment Dreaming. The predictive objective improves task progress score across all seven reported task–embodiment settings while adding no deployment-time computation.

We compute Gromov–Wasserstein (GW) distances between embodiment-specific distributions of shared-trunk latents extracted from Scoop demonstrations, and report the percentage change in GW for each embodiment pair. Negative values indicate greater geometric similarity.

Latent Geometry. Pretraining and embodiment dreaming both make the internal distance structure of embodiment-specific latents more similar for the same manipulation task.

We compare MXT+ with a coarse variant that merges modality-specific inputs and actions, each with and without cross-embodiment pretraining. The modular design consistently outperforms the coarse variant, with particularly pronounced gains on the higher-dimensional xArm7-Dex4 and G1-Sim-Dex3.

Modular Design. After pretraining, the coarse variant scores zero on Kitchen-Cleanup, whose first step requires a 3D velocity command for locomotion, a small but critical modality within G1's 51-dimensional action space.