We compare MXT+ with and without cross-embodiment pretraining against ACT and π_0.5 on 13 task–embodiment pairs. Pretraining raises MXT+'s mean normalized task progress from 0.63 to 0.80, pooling ID and OOD trials within each pair, with improvements on 11 of 13 pairs under ID evaluation and all 13 under OOD evaluation. The gains are particularly pronounced under OOD conditions, where mean task progress increases from 0.53 to 0.77, with large improvements on G1-Dex5 Insert and Z1-Tool Scoop. Among all evaluated methods, pretrained MXT+ achieves the highest average task progress and ranks first on 12 of the 13 pairs.

Main Results. Cross-embodiment pretraining improves MXT+ on 11 of 13 task–embodiment pairs under ID evaluation and on all 13 under OOD evaluation, where the gains are particularly pronounced.

We compare four pretraining mixtures: no pretraining, human-only, human plus other robots, and full cross-embodiment pretraining, which further includes the target robot's demonstrations on non-target tasks. All variants use the same target-embodiment, target-task demonstrations for downstream training, on Sort for R1Lite-Gripper, Sort and Insert for G1-Dex5 and xArm7-Dex4, and Kitchen-Bread for G1-Sim-Dex3.

Pretraining Data Composition. Every pretraining mixture improves average OOD performance over no pretraining, and the average gains grow as the mixture broadens. Human-only pretraining already improves OOD performance on five of six pairs, adding other robots raises ID performance on five of six pairs, and adding the target robot's other tasks improves OOD performance on every pair. Full cross-embodiment pretraining achieves the highest average task progress, improving over no pretraining by 0.12 on ID and 0.23 on OOD.

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

We compare MXT+ with MXT, which removes the embodiment-dreaming components (the embodiment tokenizer, dreaming expert, and EMA teachers) during both pretraining and fine-tuning. MXT+ improves average task progress from 0.63 to 0.78 across seven task–embodiment settings spanning five embodiments, with gains in all seven settings and the largest on G1-Dex5 Scoop.

Embodiment Dreaming. The embodiment-dreaming components improve task progress score in all seven reported task–embodiment settings, and deployment requires neither future-latent prediction nor EMA targets.

We compare MXT+ with a coarse variant that uses one tokenizer for all camera views, one for concatenated pose and proprioceptive streams, and one action expert for all action modalities within each embodiment, keeping the image encoders, shared trunk, embodiment dreaming, training objective, and token budgets unchanged. After pretraining, modular MXT+ matches or exceeds the coarse variant in every ID and OOD setting. Both variants benefit from pretraining on Z1-Tool, but pretraining reduces the coarse variant's performance on xArm7-Dex4 Scoop and both simulated kitchen tasks, whereas modular MXT+ improves in most settings and avoids these large drops.

Modular Design. After pretraining, the gap between modular and coarse policies is small on Z1-Tool but substantially larger on xArm7-Dex4 and G1-Sim-Dex3, which involve richer observation and action modalities. On Kitchen-Cleanup, which starts with locomotion and therefore depends on a 3D velocity command within G1-Sim-Dex3's 51-dimensional action space, the pretrained coarse variant scores zero.

To assess pretraining, we compare MXT+ with and without cross-embodiment pretraining; to assess embodiment dreaming, we compare MXT+ with MXT, both pretrained and fine-tuned. For each comparison, we compute Gromov–Wasserstein (GW) distances between embodiment-specific distributions of shared-trunk latents extracted from Scoop demonstrations, mean-pooling the right wrist-pose tokens of the trunk output at each sampled time step into one 512-dimensional vector. Each cell shows the percentage change in GW for one 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.